Search

In order to shorten detection times and improve average precision in embedded devices, a lightweight and high-accuracy model is proposed to detect passion fruit in complex environments (e.g., with backlighting, occlusion, overlap, sun, cloud, or rain). First, replacing the backbone network of YOLOv5 with a lightweight GhostNet model reduces the number of parameters and computational complexity while improving the detection speed. Second, a new feature branch is added to the backbone network and the feature fusion layer in the neck network is reconstructed to effectively combine the lower- and higher-level features, which improves the accuracy of the model while maintaining its lightweight nature. Finally, a knowledge distillation method is used to transfer knowledge from the more capable teacher model to the less capable student model, significantly improving the detection accuracy. The improved model is denoted as G-YOLO-NK. The average accuracy of the G-YOLO-NK network is 96.00%, which is 1.00% higher than that of the original YOLOv5s model. Furthermore, the model size is 7.14 MB, half that of the original model, and its real-time detection frame rate is 11.25 FPS when implemented on the Jetson Nano. The proposed model is found to outperform state-of-the-art models in terms of average precision and detection performance. The present work provides an effective model for real-time detection of passion fruit in complex orchard scenes, offering valuable technical support for the development of orchard picking robots and greatly improving the intelligence level of orchards.

In response to the increasing number of agents and changing task scenarios in multi-agent collaborative systems, existing collaborative strategies struggle to effectively adapt to new task scenarios. To address this challenge, this paper proposes a knowledge distillation method combined with a domain separation network (DSN-KD). This method leverages the well-performing policy network from a source task as the teacher model, utilizes a domain-separated neural network structure to correct the teacher model’s outputs as supervision, and guides the learning of agents in new tasks. The proposed method does not require the pre-design or training of complex state-action mappings, thereby reducing the cost of transfer. Experimental results in scenarios such as UAV surveillance and UAV cooperative target occupation, robot cooperative box pushing, UAV cooperative target strike, and multi-agent cooperative resource recovery in a particle simulation environment demonstrate that the DSN-KD transfer method effectively enhances the learning speed of new task policies and improves the proximity of the policy model to the theoretically optimal policy in practical tasks.

Rock image classification represents a challenging fine-grained image classification task characterized by subtle differences among closely related rock categories. Current contrastive learning methods prevalently utilized in fine-grained image classification restrict the model’s capacity to discern critical features contrastively from image pairs, and are typically too large for deployment on mobile devices used for in situ rock identification. In this work, we introduce an innovative and compact model generation framework anchored by the design of a Feature Positioning Comparison Network (FPCN). The FPCN facilitates interaction between feature vectors from localized regions within image pairs, capturing both shared and distinctive features. Further, it accommodates the variable scales of objects depicted in images, which correspond to differing quantities of inherent object information, directing the network’s attention to additional contextual details based on object size variability. Leveraging knowledge distillation, the architecture is streamlined, with a focus on nuanced information at activation boundaries to master the precise fine-grained decision boundaries, thereby enhancing the small model’s accuracy. Empirical evidence demonstrates that our proposed method based on FPCN improves the classification accuracy mobile lightweight models by nearly 2% while maintaining the same time and space consumption.

Most logit-based knowledge distillation methods transfer soft labels from the teacher model to the student model via Kullback–Leibler divergence based on softmax, an exponential normalization function. However, this exponential nature of softmax tends to prioritize the largest class (target class) while neglecting smaller ones (non-target classes), leading to an oversight of the non-target classes’s significance. To address this issue, we propose Non-Target-Class-Enhanced Knowledge Distillation (NTCE-KD) to amplify the role of non-target classes both in terms of magnitude and diversity. Specifically, we present a magnitude-enhanced Kullback–Leibler (MKL) divergence multi-shrinking the target class to enhance the impact of non-target classes in terms of magnitude. Additionally, to enrich the diversity of non-target classes, we introduce a diversity-based data augmentation strategy (DDA), further enhancing overall performance. Extensive experimental results on the CIFAR-100 and ImageNet-1k datasets demonstrate that non-target classes are of great significance and that our method achieves state-of-the-art performance across a wide range of teacher–student pairs.

Malicious social bots pose a serious threat to social network security by spreading false information and guiding bad opinions in social networks. The singularity and scarcity of single organization data and the high cost of labeling social bots have given rise to the construction of federated models that combine federated learning with social bot detection. In this paper, we first combine the federated learning framework with the Relational Graph Convolutional Neural Network (RGCN) model to achieve federated social bot detection. A class-level cross entropy loss function is applied in the local model training to mitigate the effects of the class imbalance problem in local data. To address the data heterogeneity issue from multiple participants, we optimize the classical federated learning algorithm by applying knowledge distillation methods. Specifically, we adjust the client-side and server-side models separately: training a global generator to generate pseudo-samples based on the local data distribution knowledge to correct the optimization direction of client-side classification models, and integrating client-side classification models’ knowledge on the server side to guide the training of the global classification model. We conduct extensive experiments on widely used datasets, and the results demonstrate the effectiveness of our approach in social bot detection in heterogeneous data scenarios. Compared to baseline methods, our approach achieves a nearly 3–10% improvement in detection accuracy when the data heterogeneity is larger. Additionally, our method achieves the specified accuracy with minimal communication rounds.

The purpose of this paper is to propose a novel transfer learning regularization method based on knowledge distillation. Recently, transfer learning methods have been used in various fields. However, problems such as knowledge loss still occur during the process of transfer learning to a new target dataset. To solve these problems, there are various regularization methods based on knowledge distillation techniques. In this paper, we propose a transfer learning regularization method based on feature map alignment used in the field of knowledge distillation. The proposed method is composed of two attention-based submodules: self-pixel attention (SPA) and global channel attention (GCA). The self-pixel attention submodule utilizes both the feature maps of the source and target models, so that it provides an opportunity to jointly consider the features of the target and the knowledge of the source. The global channel attention submodule determines the importance of channels through all layers, unlike the existing methods that calculate these only within a single layer. Accordingly, transfer learning regularization is performed by considering both the interior of each single layer and the depth of the entire layer. Consequently, the proposed method using both of these submodules showed overall improved classification accuracy than the existing methods in classification experiments on commonly used datasets.

The ongoing biodiversity crisis, driven by factors such as land-use change and global warming, emphasizes the need for effective ecological monitoring methods. Acoustic monitoring of biodiversity has emerged as an important monitoring tool. Detecting human voices in soundscape monitoring projects is useful both for analyzing human disturbance and for privacy filtering. Despite significant strides in deep learning in recent years, the deployment of large neural networks on compact devices poses challenges due to memory and latency constraints. Our approach focuses on leveraging knowledge distillation techniques to design efficient, lightweight student models for speech detection in bioacoustics. In particular, we employed the MobileNetV3-Small-Pi model to create compact yet effective student architectures to compare against the larger EcoVAD teacher model, a well-regarded voice detection architecture in eco-acoustic monitoring. The comparative analysis included examining various configurations of the MobileNetV3-Small-Pi-derived student models to identify optimal performance. Additionally, a thorough evaluation of different distillation techniques was conducted to ascertain the most effective method for model selection. Our findings revealed that the distilled models exhibited comparable performance to the EcoVAD teacher model, indicating a promising approach to overcoming computational barriers for real-time ecological monitoring.

Compared to fault diagnosis across operating conditions, the differences in data distribution between devices are more pronounced and better aligned with practical application needs. However, current research on transfer learning inadequately addresses fault diagnosis issues across devices. To better balance the relationship between computational resources and diagnostic accuracy, a knowledge distillation-based lightweight transfer learning framework for rolling bearing diagnosis is proposed in this study. Specifically, a deep teacher–student model based on variable-scale residual networks is constructed to learn domain-invariant features relevant to fault classification within both the source and target domain data. Subsequently, a knowledge distillation framework incorporating a temperature factor is established to transfer fault features learned by the large teacher model in the source domain to the smaller student model, thereby reducing computational and parameter overhead. Finally, a multi-kernel domain adaptation method is employed to capture the feature probability distribution distance of fault characteristics between the source and target domains in Reproducing Kernel Hilbert Space (RKHS), and domain-invariant features are learned by minimizing the distribution distance between them. The effectiveness and applicability of the proposed method in situations of incomplete data across device types were validated through two engineering cases, spanning device models and transitioning from laboratory equipment to real-world operational devices.

In recent computer vision research, the pursuit of improved classification performance often leads to the adoption of complex, large-scale models. However, the actual deployment of such extensive models poses significant challenges in environments constrained by limited computing power and storage capacity. Consequently, this study is dedicated to addressing these challenges by focusing on innovative methods that enhance the classification performance of lightweight models. We propose a novel method to compress the knowledge learned by a large model into a lightweight one so that the latter can also achieve good performance in few-shot classification tasks. Specifically, we propose a dual-faceted knowledge distillation strategy that combines output-based and intermediate feature-based methods. The output-based method concentrates on distilling knowledge related to base class labels, while the intermediate feature-based approach, augmented by feature error distribution calibration, tackles the potential non-Gaussian nature of feature deviations, thereby boosting the effectiveness of knowledge transfer. Experiments conducted on MiniImageNet, CIFAR-FS, and CUB datasets demonstrate the superior performance of our method over state-of-the-art lightweight models, particularly in five-way one-shot and five-way five-shot tasks.

An equalizer based on a recurrent neural network (RNN), especially with a bidirectional gated recurrent unit (biGRU) structure, is a good choice to deal with nonlinear damage and inter-symbol interference (ISI) in optical communication systems because of its excellent performance in processing time series information. However, its recursive structure prevents the parallelization of the computation, resulting in a low equalization rate. In order to improve the speed without compromising the equalization performance, we propose a minimalist 1D convolutional neural network (CNN) equalizer, which is reconverted from a biGRU with knowledge distillation (KD). In this work, we applied KD to regression problems and explain how KD helps students learn from teachers in solving regression problems. In addition, we compared the biGRU, 1D-CNN after KD and 1D-CNN without KD in terms of Q-factor and equalization velocity. The experimental data showed that the Q-factor of the 1D-CNN increased by 1 dB after KD learning from the biGRU, and KD increased the RoP sensitivity of the 1D-CNN by 0.89 dB with the HD-FEC threshold of 1 × 10−3. At the same time, compared with the biGRU, the proposed 1D-CNN equalizer reduced the computational time consumption by 97% and the number of trainable parameters by 99.3%, with only a 0.5 dB Q-factor penalty. The results demonstrate that the proposed minimalist 1D-CNN equalizer holds significant promise for future practical deployments in optical wireless communication systems.

The identification and classification of traditional Chinese herbal medicines demand significant time and expertise. We propose the dual-teacher supervised decay (DTSD) approach, an enhancement for Chinese herbal medicine recognition utilizing a refined knowledge distillation model. The DTSD method refines output soft labels, adapts attenuation parameters, and employs a dynamic combination loss in the teacher model. Implemented on the lightweight MobileNet_v3 network, the methodology is deployed successfully in a mobile application. Experimental results reveal that incorporating the exponential warmup learning rate reduction strategy during training optimizes the knowledge distillation model, achieving an average classification accuracy of 98.60% for 10 types of Chinese herbal medicine images. The model boasts an average detection time of 0.0172 s per image, with a compressed size of 10 MB. Comparative experiments demonstrate the superior performance of our refined model over DenseNet121, ResNet50_vd, Xception65, and EfficientNetB1. This refined model not only introduces an approach to Chinese herbal medicine image recognition but also provides a practical solution for lightweight models in mobile applications.

Monocular depth estimation is a task aimed at predicting pixel-level distances from a single RGB image. This task holds significance in various applications including autonomous driving and robotics. In particular, the recognition of surrounding environments is important to avoid collisions during autonomous parking. Fisheye cameras are adequate to acquire visual information from a wide field of view, reducing blind spots and preventing potential collisions. While there have been increasing demands for fisheye cameras in visual-recognition systems, existing research on depth estimation has primarily focused on pinhole camera images. Moreover, depth estimation from fisheye images poses additional challenges due to strong distortion and the lack of public datasets. In this work, we propose a novel underground parking lot dataset called JBNU-Depth360, which consists of fisheye camera images and their corresponding LiDAR projections. Our proposed dataset was composed of 4221 pairs of fisheye images and their corresponding LiDAR point clouds, which were obtained from six driving sequences. Furthermore, we employed a knowledge-distillation technique to improve the performance of the state-of-the-art depth-estimation models. The teacher–student learning framework allows the neural network to leverage the information in dense depth predictions and sparse LiDAR projections. Experiments were conducted on the KITTI-360 and JBNU-Depth360 datasets for analyzing the performance of existing depth-estimation models on fisheye camera images. By utilizing the self-distillation technique, the AbsRel and SILog error metrics were reduced by 1.81% and 1.55% on the JBNU-Depth360 dataset. The experimental results demonstrated that the self-distillation technique is beneficial to improve the performance of depth-estimation models.

In this study, we propose a knowledge distillation (KD) method for segmenting off-road environment range images. Unlike urban environments, off-road terrains are irregular and pose a higher risk to hardware. Therefore, off-road self-driving systems are required to be computationally efficient. We used LiDAR point cloud range images to address this challenge. The three-dimensional (3D) point cloud data, which are rich in detail, require substantial computational resources. To mitigate this problem, we employ a projection method to convert the image into a two-dimensional (2D) image format using depth information. Our soft label-based knowledge distillation (SLKD) effectively transfers knowledge from a large teacher network to a lightweight student network. We evaluated SLKD using the RELLIS-3D off-road environment dataset, measuring the performance with respect to the mean intersection of union (mIoU) and GPU floating point operations per second (GFLOPS). The experimental results demonstrate that SLKD achieves a favorable trade-off between mIoU and GFLOPS when comparing teacher and student networks. This approach shows promise for enabling efficient off-road autonomous systems with reduced computational costs.

Hyperspectral images provide a wealth of spectral and spatial information, offering significant advantages for the purpose of tracking objects. However, Siamese trackers are unable to fully exploit spectral features due to the limited number of hyperspectral videos. The high-dimensional nature of hyperspectral images complicates the model training process. In order to address the aforementioned issues, this article proposes a hyperspectral object tracking (HOT) algorithm callled SiamPKHT, which leverages the SiamCAR model by incorporating pyramid shuffle attention (PSA) and knowledge distillation (KD). First, the PSA module employs pyramid convolutions to extract multiscale features. In addition, shuffle attention is adopted to capture relationships between different channels and spatial positions, thereby obtaining good features with a stronger classification performance. Second, KD is introduced under the guidance of a pre-trained RGB tracking model, which deals with the problem of overfitting in HOT. Experiments using HOT2022 data indicate that the designed SiamPKHT achieves better performance compared to the baseline method (SiamCAR) and other state-of-the-art HOT algorithms. It also achieves real-time requirements at 43 frames per second.

In the target-controlled infusion (TCI) of propofol and remifentanil intravenous anesthesia, accurate prediction of the depth of anesthesia (DOA) is very challenging. Patients with different physiological characteristics have inconsistent pharmacodynamic responses during different stages of anesthesia. For example, in TCI, older adults transition smoothly from the induction period to the maintenance period, while younger adults are more prone to anesthetic awareness, resulting in different DOA data distributions among patients. To address these problems, a deep learning framework that incorporates domain adaptation and knowledge distillation and uses propofol and remifentanil doses at historical moments to continuously predict the bispectral index (BIS) is proposed in this paper. Specifically, a modified adaptive recurrent neural network (AdaRNN) is adopted to address data distribution differences among patients. Moreover, a knowledge distillation pipeline is developed to train the prediction network by enabling it to learn intermediate feature representations of the teacher network. The experimental results show that our method exhibits better performance than existing approaches during all anesthetic phases in the TCI of propofol and remifentanil intravenous anesthesia. In particular, our method outperforms some state-of-the-art methods in terms of root mean square error and mean absolute error by 1 and 0.8, respectively, in the internal dataset as well as in the publicly available dataset.

Image anomaly detection is a trending research topic in computer vision. The objective is to build models using available normal samples to detect various abnormal images without depending on real abnormal samples. It has high research significance and value for applications in the detection of defects in product appearance, medical image analysis, hyperspectral image processing, and other fields. This paper proposes an image anomaly detection algorithm based on feature distillation and an autoencoder structure, which uses the feature distillation structure of a dual-teacher network to train the encoder, thus suppressing the reconstruction of abnormal regions. This system also introduces an attention mechanism to highlight the detection objects, achieving effective detection of different defects in product appearance. In addition, this paper proposes a method for anomaly evaluation based on patch similarity that calculates the difference between the reconstructed image and the input image according to different regions of the image, thus improving the sensitivity and accuracy of the anomaly score. This paper conducts experiments on several datasets, and the results show that the proposed algorithm has superior performance in image anomaly detection. It achieves 98.8% average AUC on the SMDC-DET dataset and 98.9% average AUC on the MVTec-AD dataset.

Deep-sea biological detection is essential for deep-sea resource research and conservation. However, due to the poor image quality and insufficient image samples in the complex deep-sea imaging environment, resulting in poor detection results. Furthermore, most existing detection models accomplish high precision at the expense of increased complexity, and leading cannot be well deployed in the deep-sea environment. To alleviate these problems, a detection method for deep-sea organisms based on lightweight YOLOv5n is proposed. First, a lightweight YOLOv5n is created. The proposed image enhancement method based on global and local contrast fusion (GLCF) is introduced into the input layer of YOLOv5n to address the problem of color deviation and low contrast in the image. At the same time, a Bottleneck based on the Ghost module and simAM (GS-Bottleneck) is developed to achieve a lightweight model while ensuring sure detection performance. Second, a transfer learning strategy combined with knowledge distillation (TLKD) is designed, which can reduce the dependence of the model on the amount of data and improve the generalization ability to enhance detection accuracy. Experimental results on the deep-sea biological dataset show that the proposed method achieves good detection accuracy and speed, outperforming existing methods.

Continual semantic segmentation (CSS) aims to learn new tasks sequentially and extract object(s) and stuff represented by pixel-level maps of new categories while preserving the original segmentation capabilities even when the old class data is absent. Current CSS methods typically preserve the capacities of segmenting old classes via knowledge distillation, which encounters the limitations of insufficient utilization of the semantic knowledge, i.e., only distilling the last layer of the feature encoder, and the semantic shift of background caused by directly distilling the entire feature map of the decoder. In this paper, we propose a novel CCS method based on scale-hybrid distillation and knowledge disentangling to address these limitations. Firstly, we propose a scale-hybrid group semantic distillation (SGD) method for encoding, which transfers the multi-scale knowledge from the old model’s feature encoder with group pooling refinement to improve the stability of new models. Then, the knowledge disentangling distillation (KDD) method for decoding is proposed to distillate feature maps with the guidance of the old class regions and reduce incorrect guides from old models towards better plasticity. Extensive experiments are conducted on the Pascal VOC and ADE20K datasets. Competitive performance compared with other state-of-the-art methods demonstrates the effectiveness of our proposed method.

Knowledge distillation (KD) is a well-established technique for compressing neural networks and has gained increasing attention in object detection tasks. However, typical object detection distillation methods use fixed-level semantic features for distillation, which might not be best for all training stages and samples. In this paper, a multilayer semantic feature adaptive distillation (MSFAD) method is proposed that uses a routing network composed of a teacher and a student detector, along with an agent network for decision making. Specifically, the inputs to the proxy network consist of the features output by the neck structures of the teacher and student detectors, and the output is a decision on which features to choose for distillation. The MSFAD method improves the distillation training process by enabling the student detector to automatically select valuable semantic-level features from the teacher detector. Experimental results demonstrated that the proposed method increased the mAP50 of YOLOv5s by 3.4% and the mAP50–90 by 3.3%. Additionally, YOLOv5n with only 1.9 M parameters achieved detection performance comparable to that of YOLOv5s.

As the development of the Internet of Things (IoT) continues, Federated Learning (FL) is gaining popularity as a distributed machine learning framework that does not compromise the data privacy of each participant. However, the data held by enterprises and factories in the IoT often have different distribution properties (Non-IID), leading to poor results in their federated learning. This problem causes clients to forget about global knowledge during their local training phase and then tends to slow convergence and degrades accuracy. In this work, we propose a method named FedRAD, which is based on relational knowledge distillation that further enhances the mining of high-quality global knowledge by local models from a higher-dimensional perspective during their local training phase to better retain global knowledge and avoid forgetting. At the same time, we devise an entropy-wise adaptive weights module (EWAW) to better regulate the proportion of loss in single-sample knowledge distillation versus relational knowledge distillation so that students can weigh losses based on predicted entropy and learn global knowledge more effectively. A series of experiments on CIFAR10 and CIFAR100 show that FedRAD has better performance in terms of convergence speed and classification accuracy compared to other advanced FL methods.

Safety helmets are essential in various indoor and outdoor workplaces, such as metallurgical high-temperature operations and high-rise building construction, to avoid injuries and ensure safety in production. However, manual supervision is costly and prone to lack of enforcement and interference from other human factors. Moreover, small target object detection frequently lacks precision. Improving safety helmets based on the helmet detection algorithm can address these issues and is a promising approach. In this study, we proposed a modified version of the YOLOv5s network, a lightweight deep learning-based object identification network model. The proposed model extends the YOLOv5s network model and enhances its performance by recalculating the prediction frames, utilizing the IoU metric for clustering, and modifying the anchor frames with the K-means++ method. The global attention mechanism (GAM) and the convolutional block attention module (CBAM) were added to the YOLOv5s network to improve its backbone and neck networks. By minimizing information feature loss and enhancing the representation of global interactions, these attention processes enhance deep learning neural networks’ capacity for feature extraction. Furthermore, the CBAM is integrated into the CSP module to improve target feature extraction while minimizing computation for model operation. In order to significantly increase the efficiency and precision of the prediction box regression, the proposed model additionally makes use of the most recent SIoU (SCYLLA-IoU LOSS) as the bounding box loss function. Based on the improved YOLOv5s model, knowledge distillation technology is leveraged to realize the light weight of the network model, thereby reducing the computational workload of the model and improving the detection speed to meet the needs of real-time monitoring. The experimental results demonstrate that the proposed model outperforms the original YOLOv5s network model in terms of accuracy (Precision), recall rate (Recall), and mean average precision (mAP). The proposed model may more effectively identify helmet use in low-light situations and at a variety of distances.

In this paper, we addressed the challenges in sorting high-yield apple cultivars that traditionally relied on manual labor or system-based defect detection. Existing single-camera methods failed to uniformly capture the entire surface of apples, potentially leading to misclassification due to defects in unscanned areas. Various methods were proposed where apples were rotated using rollers on a conveyor. However, since the rotation was highly random, it was difficult to scan the apples uniformly for accurate classification. To overcome these limitations, we proposed a multi-camera-based apple sorting system with a rotation mechanism that ensured uniform and accurate surface imaging. The proposed system applied a rotation mechanism to individual apples while simultaneously utilizing three cameras to capture the entire surface of the apples. This method offered the advantage of quickly and uniformly acquiring the entire surface compared to single-camera and random rotation conveyor setups. The images captured by the system were analyzed using a CNN classifier deployed on embedded hardware. To maintain excellent CNN classifier performance while reducing its size and inference time, we employed knowledge distillation techniques. The CNN classifier demonstrated an inference speed of 0.069 s and an accuracy of 93.83% based on 300 apple samples. The integrated system, which included the proposed rotation mechanism and multi-camera setup, took a total of 2.84 s to sort one apple. Our proposed system provided an efficient and precise solution for detecting defects on the entire surface of apples, improving the sorting process with high reliability.

Three-dimensional (3D) real-time object detection and tracking is an important task in the case of autonomous vehicles and road and railway smart mobility, in order to allow them to analyze their environment for navigation and obstacle avoidance purposes. In this paper, we improve the efficiency of 3D monocular object detection by using dataset combination and knowledge distillation, and by creating a lightweight model. Firstly, we combine real and synthetic datasets to increase the diversity and richness of the training data. Then, we use knowledge distillation to transfer the knowledge from a large, pre-trained model to a smaller, lightweight model. Finally, we create a lightweight model by selecting the combinations of width, depth & resolution in order to reach a target complexity and computation time. Our experiments showed that using each method improves either the accuracy or the efficiency of our model with no significant drawbacks. Using all these approaches is especially useful for resource-constrained environments, such as self-driving cars and railway systems.

Recent advances with large-scale pre-trained language models (e.g., BERT) have brought significant potential to natural language processing. However, the large model size hinders their use in IoT and edge devices. Several studies have utilized task-specific knowledge distillation to compress the pre-trained language models. However, to reduce the number of layers in a large model, a sound strategy for distilling knowledge to a student model with fewer layers than the teacher model is lacking. In this work, we present Layer-wise Adaptive Distillation (LAD), a task-specific distillation framework that can be used to reduce the model size of BERT. We design an iterative aggregation mechanism with multiple gate blocks in LAD to adaptively distill layer-wise internal knowledge from the teacher model to the student model. The proposed method enables an effective knowledge transfer process for a student model, without skipping any teacher layers. The experimental results show that both the six-layer and four-layer LAD student models outperform previous task-specific distillation approaches during GLUE tasks.

Face alignment is widely used in high-level face analysis applications, such as human activity recognition and human–computer interaction. However, most existing models involve a large number of parameters and are computationally inefficient in practical applications. In this paper, we aim to build a lightweight facial landmark detector by proposing a network-level architecture-slimming method. Concretely, we introduce a selective feature fusion mechanism to quantify and prune redundant transformation and aggregation operations in a high-resolution supernetwork. Moreover, we develop a triple knowledge distillation scheme to further refine a slimmed network, where two peer student networks could learn the implicit landmark distributions from each other while absorbing the knowledge from a teacher network. Extensive experiments on challenging benchmarks, including 300W, COFW, and WFLW, demonstrate that our approach achieves competitive performance with a better trade-off between the number of parameters (0.98 M–1.32 M) and the number of floating-point operations (0.59 G–0.6 G) when compared to recent state-of-the-art methods.

Sentiment classification is a key task in exploring people’s opinions; improved sentiment classification can help individuals make better decisions. Social media users are increasingly using both images and text to express their opinions and share their experiences, instead of only using text in conventional social media. As a result, understanding how to fully utilize them is critical in a variety of activities, including sentiment classification. In this work, we provide a fresh multimodal sentiment classification approach: visual distillation and attention network or VisdaNet. First, this method proposes a knowledge augmentation module, which overcomes the lack of information in short text by integrating the information of image captions and short text; secondly, aimed at the information control problem in the multi-modal fusion process in the product review scene, this paper proposes a knowledge distillation based on the CLIP module to reduce the noise information of the original modalities and improve the quality of the original modal information. Finally, regarding the single-text multi-image fusion problem in the product review scene, this paper proposes visual aspect attention based on the CLIP module, which correctly models the text-image interaction relationship in special scenes and realizes feature-level fusion across modalities. The results of the experiment on the Yelp multimodal dataset reveal that our model outperforms the previous SOTA model. Furthermore, the ablation experiment results demonstrate the efficacy of various tactics in the suggested model.

Deep learning-based Human Activity Recognition (HAR) systems received a lot of interest for health monitoring and activity tracking on wearable devices. The availability of large and representative datasets is often a requirement for training accurate deep learning models. To keep private data on users’ devices while utilizing them to train deep learning models on huge datasets, Federated Learning (FL) was introduced as an inherently private distributed training paradigm. However, standard FL (FedAvg) lacks the capability to train heterogeneous model architectures. In this paper, we propose Federated Learning via Augmented Knowledge Distillation (FedAKD) for distributed training of heterogeneous models. FedAKD is evaluated on two HAR datasets: A waist-mounted tabular HAR dataset and a wrist-mounted time-series HAR dataset. FedAKD is more flexible than standard federated learning (FedAvg) as it enables collaborative heterogeneous deep learning models with various learning capacities. In the considered FL experiments, the communication overhead under FedAKD is 200X less compared with FL methods that communicate models’ gradients/weights. Relative to other model-agnostic FL methods, results show that FedAKD boosts performance gains of clients by up to 20 percent. Furthermore, FedAKD is shown to be relatively more robust under statistical heterogeneous scenarios.

This paper proposes a photoelectric target detection algorithm for NVIDIA Jeston Nano embedded devices, exploiting the characteristics of active and passive differential images of lasers after denoising. An adaptive threshold segmentation method was developed based on the statistical characteristics of photoelectric target echo light intensity, which effectively improves detection of the target area. The proposed method’s effectiveness is compared and analyzed against a typical lightweight network that was knowledge-distilled by ResNet18 on target region detection tasks. Furthermore, TensorRT technology was applied to accelerate inference and deploy on hardware platforms the lightweight network Shuffv2_x0_5. The experimental results demonstrate that the developed method’s accuracy rate reaches 97.15%, the false alarm rate is 4.87%, and the detection rate can reach 29 frames per second for an image resolution of 640 × 480 pixels.

Depth perception capability is one of the essential requirements for various autonomous driving platforms. However, accurate depth estimation in a real-world setting is still a challenging problem due to high computational costs. In this paper, we propose a lightweight depth completion network for depth perception in real-world environments. To effectively transfer a teacher’s knowledge, useful for the depth completion, we introduce local similarity-preserving knowledge distillation (LSPKD), which allows similarities between local neighbors to be transferred during the distillation. With our LSPKD, a lightweight student network is precisely guided by a heavy teacher network, regardless of the density of the ground-truth data. Experimental results demonstrate that our method is effective to reduce computational costs during both training and inference stages while achieving superior performance over other lightweight networks.

The goal of gaze estimation is to estimate a gaze vector from an image containing a face or eye(s). Most existing studies use pre-defined fixed-resolution images to estimate the gaze vector. However, images captured from in-the-wild environments may have various resolutions, and variation in resolution can degrade gaze estimation performance. To address this problem, a gaze estimation method from arbitrary-sized low-resolution images is proposed. The basic idea of the proposed method is to combine knowledge distillation and feature adaptation. Knowledge distillation helps the gaze estimator for arbitrary-sized images generate a feature map similar to that from a high-resolution image. Feature adaptation makes creating a feature map adaptive to various resolutions of an input image possible by using a low-resolution image and its scale information together. It is shown that combining these two ideas improves gaze estimation performance substantially in the ablation study. It is also demonstrated that the proposed method can be generalized to other popularly used gaze estimation models through experiments using various backbones.

Current distillation methods only distill between corresponding layers, and do not consider the knowledge contained in preceding layers. To solve this problem, we analyzed the guiding effect of the inferior features of a teacher model on the coordinate feature of a student model, and proposed inferior and coordinate distillation for object detectors. The proposed method utilizes the rich information contained in different layers of the teacher model; such that the student model can review the old information and learn the new information, in addition to the dark knowledge in the teacher model. Moreover, the refine module is used to align the features of different layers, distinguish the spatial and channel to extract attention, strengthen the correlation between the features of different stages, and prevent the disorder caused by merging. Exclusive experiments were conducted on different object detectors. The results for the mean average precision (mAP) obtained using Faster R-CNN, RetinaNet, and fully convolutional one-stage object detector (FCOS) with ResNet-50 as its backbone were 40.5%, 39.8%, and 42.8% with regard to the COCO dataset, respectively; which are 2.1%, 2.4%, and 4.3% higher than the benchmark, respectively.

While recent deep learning-based stereo-matching networks have shown outstanding advances, there are still some unsolved challenges. First, most state-of-the-art stereo models employ 3D convolutions for 4D cost volume aggregation, which limit the deployment of networks for resource-limited mobile environments owing to heavy consumption of computation and memory. Although there are some efficient networks, most of them still require a heavy computational cost to incorporate them to mobile computing devices in real-time. Second, most stereo networks indirectly supervise cost volumes through disparity regression loss by using the softargmax function. This causes problems in ambiguous regions, such as the boundaries of objects, because there are many possibilities for unreasonable cost distributions which result in overfitting problem. A few works deal with this problem by generating artificial cost distribution using only the ground truth disparity value that is insufficient to fully regularize the cost volume. To address these problems, we first propose an efficient multi-scale sequential feature fusion network (MSFFNet). Specifically, we connect multi-scale SFF modules in parallel with a cross-scale fusion function to generate a set of cost volumes with different scales. These cost volumes are then effectively combined using the proposed interlaced concatenation method. Second, we propose an adaptive cost-volume-filtering (ACVF) loss function that directly supervises our estimated cost volume. The proposed ACVF loss directly adds constraints to the cost volume using the probability distribution generated from the ground truth disparity map and that estimated from the teacher network which achieves higher accuracy. Results of several experiments using representative datasets for stereo matching show that our proposed method is more efficient than previous methods. Our network architecture consumes fewer parameters and generates reasonable disparity maps with faster speed compared with the existing state-of-the art stereo models. Concretely, our network achieves 1.01 EPE with runtime of 42 ms, 2.92 M parameters, and 97.96 G FLOPs on the Scene Flow test set. Compared with PSMNet, our method is 89% faster and 7% more accurate with 45% fewer parameters.

This work addresses the challenge of building an accurate and generalizable periocular recognition model with a small number of learnable parameters. Deeper (larger) models are typically more capable of learning complex information. For this reason, knowledge distillation (kd) was previously proposed to carry this knowledge from a large model (teacher) into a small model (student). Conventional KD optimizes the student output to be similar to the teacher output (commonly classification output). In biometrics, comparison (verification) and storage operations are conducted on biometric templates, extracted from pre-classification layers. In this work, we propose a novel template-driven KD approach that optimizes the distillation process so that the student model learns to produce templates similar to those produced by the teacher model. We demonstrate our approach on intra- and cross-device periocular verification. Our results demonstrate the superiority of our proposed approach over a network trained without KD and networks trained with conventional (vanilla) KD. For example, the targeted small model achieved an equal error rate (EER) value of 22.2% on cross-device verification without KD. The same model achieved an EER of 21.9% with the conventional KD, and only 14.7% EER when using our proposed template-driven KD.

Multi-person pose estimation has been gaining considerable interest due to its use in several real-world applications, such as activity recognition, motion capture, and augmented reality. Although the improvement of the accuracy and speed of multi-person pose estimation techniques has been recently studied, limitations still exist in balancing these two aspects. In this paper, a novel knowledge distilled lightweight top-down pose network (KDLPN) is proposed that balances computational complexity and accuracy. For the first time in multi-person pose estimation, a network that reduces computational complexity by applying a “Pelee” structure and shuffles pixels in the dense upsampling convolution layer to reduce the number of channels is presented. Furthermore, to prevent performance degradation because of the reduced computational complexity, knowledge distillation is applied to establish the pose estimation network as a teacher network. The method performance is evaluated on the MSCOCO dataset. Experimental results demonstrate that our KDLPN network significantly reduces 95% of the parameters required by state-of-the-art methods with minimal performance degradation. Moreover, our method is compared with other pose estimation methods to substantiate the importance of computational complexity reduction and its effectiveness.

Human-robot interaction has received a lot of attention as collaborative robots became widely utilized in many industrial fields. Among techniques for human-robot interaction, collision identification is an indispensable element in collaborative robots to prevent fatal accidents. This paper proposes a deep learning method for identifying external collisions in 6-DoF articulated robots. The proposed method expands the idea of CollisionNet, which was previously proposed for collision detection, to identify the locations of external forces. The key contribution of this paper is uncertainty-aware knowledge distillation for improving the accuracy of a deep neural network. Sample-level uncertainties are estimated from a teacher network, and larger penalties are imposed for uncertain samples during the training of a student network. Experiments demonstrate that the proposed method is effective for improving the performance of collision identification.

Deep neural networks have achieved state-of-the-art performance in image classification. Due to this success, deep learning is now also being applied to other data modalities such as multispectral images, lidar and radar data. However, successfully training a deep neural network requires a large reddataset. Therefore, transitioning to a new sensor modality (e.g., from regular camera images to multispectral camera images) might result in a drop in performance, due to the limited availability of data in the new modality. This might hinder the adoption rate and time to market for new sensor technologies. In this paper, we present an approach to leverage the knowledge of a teacher network, that was trained using the original data modality, to improve the performance of a student network on a new data modality: a technique known in literature as knowledge distillation. By applying knowledge distillation to the problem of sensor transition, we can greatly speed up this process. We validate this approach using a multimodal version of the MNIST dataset. Especially when little data is available in the new modality (i.e., 10 images), training with additional teacher supervision results in increased performance, with the student network scoring a test set accuracy of 0.77, compared to an accuracy of 0.37 for the baseline. We also explore two extensions to the default method of knowledge distillation, which we evaluate on a multimodal version of the CIFAR-10 dataset: an annealing scheme for the hyperparameter α and selective knowledge distillation. Of these two, the first yields the best results. Choosing the optimal annealing scheme results in an increase in test set accuracy of 6%. Finally, we apply our method to the real-world use case of skin lesion classification.

Explaining the prediction of deep neural networks makes the networks more understandable and trusted, leading to their use in various mission critical tasks. Recent progress in the learning capability of networks has primarily been due to the enormous number of model parameters, so that it is usually hard to interpret their operations, as opposed to classical white-box models. For this purpose, generating saliency maps is a popular approach to identify the important input features used for the model prediction. Existing explanation methods typically only use the output of the last convolution layer of the model to generate a saliency map, lacking the information included in intermediate layers. Thus, the corresponding explanations are coarse and result in limited accuracy. Although the accuracy can be improved by iteratively developing a saliency map, this is too time-consuming and is thus impractical. To address these problems, we proposed a novel approach to explain the model prediction by developing an attentive surrogate network using the knowledge distillation. The surrogate network aims to generate a fine-grained saliency map corresponding to the model prediction using meaningful regional information presented over all network layers. Experiments demonstrated that the saliency maps are the result of spatially attentive features learned from the distillation. Thus, they are useful for fine-grained classification tasks. Moreover, the proposed method runs at the rate of 24.3 frames per second, which is much faster than the existing methods by orders of magnitude.

While growing instruments generate more and more airborne or satellite images, the bottleneck in remote sensing (RS) scene classification has shifted from data limits toward a lack of ground truth samples. There are still many challenges when we are facing unknown environments, especially those with insufficient training data. Few-shot classification offers a different picture under the umbrella of meta-learning: digging rich knowledge from a few data are possible. In this work, we propose a method named RS-SSKD for few-shot RS scene classification from a perspective of generating powerful representation for the downstream meta-learner. Firstly, we propose a novel two-branch network that takes three pairs of original-transformed images as inputs and incorporates Class Activation Maps (CAMs) to drive the network mining, the most relevant category-specific region. This strategy ensures that the network generates discriminative embeddings. Secondly, we set a round of self-knowledge distillation to prevent overfitting and boost the performance. Our experiments show that the proposed method surpasses current state-of-the-art approaches on two challenging RS scene datasets: NWPU-RESISC45 and RSD46-WHU. Finally, we conduct various ablation experiments to investigate the effect of each component of the proposed method and analyze the training time of state-of-the-art methods and ours.

In this paper, we propose an efficient knowledge distillation method to train light networks using heavy networks for semantic segmentation. Most semantic segmentation networks that exhibit good accuracy are based on computationally expensive networks. These networks are not suitable for mobile applications using vision sensors, because computational resources are limited in these environments. In this view, knowledge distillation, which transfers knowledge from heavy networks acting as teachers to light networks as students, is suitable methodology. Although previous knowledge distillation approaches have been proven to improve the performance of student networks, most methods have some limitations. First, they tend to use only the spatial correlation of feature maps and ignore the relational information of their channels. Second, they can transfer false knowledge when the results of the teacher networks are not perfect. To address these two problems, we propose two loss functions: a channel and spatial correlation (CSC) loss function and an adaptive cross entropy (ACE) loss function. The former computes the full relationship of both the channel and spatial information in the feature map, and the latter adaptively exploits one-hot encodings using the ground truth labels and the probability maps predicted by the teacher network. To evaluate our method, we conduct experiments on scene parsing datasets: Cityscapes and Camvid. Our method presents significantly better performance than previous methods.

As artificial intelligence (AI)- or deep-learning-based technologies become more popular, the main research interest in the field is not only on their accuracy, but also their efficiency, e.g., the ability to give immediate results on the users’ inputs. To achieve this, there have been many attempts to embed deep learning technology on intelligent sensors. However, there are still many obstacles in embedding a deep network in sensors with limited resources. Most importantly, there is an apparent trade-off between the complexity of a network and its processing time, and finding a structure with a better trade-off curve is vital for successful applications in intelligent sensors. In this paper, we propose two strategies for designing a compact deep network that maintains the required level of performance even after minimizing the computations. The first strategy is to automatically determine the number of parameters of a network by utilizing group sparsity and knowledge distillation (KD) in the training process. By doing so, KD can compensate for the possible losses in accuracy caused by enforcing sparsity. Nevertheless, a problem in applying the first strategy is the unclarity in determining the balance between the accuracy improvement due to KD and the parameter reduction by sparse regularization. To handle this balancing problem, we propose a second strategy: a feedback control mechanism based on the proportional control theory. The feedback control logic determines the amount of emphasis to be put on network sparsity during training and is controlled based on the comparative accuracy losses of the teacher and student models in the training. A surprising fact here is that this control scheme not only determines an appropriate trade-off point, but also improves the trade-off curve itself. The results of experiments on CIFAR-10, CIFAR-100, and ImageNet32 × 32 datasets show that the proposed method is effective in building a compact network while preventing performance degradation due to sparsity regularization much better than other baselines.

We investigated a machine-learning-based fast banknote serial number recognition method. Unlike existing methods, the proposed method not only recognizes multi-digit serial numbers simultaneously but also detects the region of interest for the serial number automatically from the input image. Furthermore, the proposed method uses knowledge distillation to compress a cumbersome deep-learning model into a simple model to achieve faster computation. To automatically decide hyperparameters for knowledge distillation, we applied the Bayesian optimization method. In experiments using Japanese Yen, Korean Won, and Euro banknotes, the proposed method showed significant improvement in computation time while maintaining a performance comparable to a sequential region of interest (ROI) detection and classification method.