Search

Waterfowl populations monitoring is essential for wetland conservation. Lately, deep learning techniques have shown promising advancements in detecting waterfowl in aerial images. In this paper, we present performance evaluation of several popular supervised and semi-supervised deep learning models for waterfowl detection in aerial images using four new image datasets containing 197,642 annotations. The best-performing model, Faster R-CNN, achieved 95.38% accuracy in terms of mAP. Semi-supervised learning models outperformed supervised models when the same amount of labeled data was used for training. Additionally, we present performance evaluation of several deep learning models on waterfowl classifications on aerial images using a new real-bird classification dataset consisting of 6,986 examples and a new decoy classification dataset consisting of about 10,000 examples per category of 20 categories. The best model achieved accuracy of 91.58% on the decoy dataset and 82.88% on the real-bird dataset. [ABSTRACT FROM AUTHOR]

Background: Breast cancer is arguably one of the leading causes of death among women around the world. The automation of the early detection process and classification of breast masses has been a prominent focus for researchers in the past decade. The utilization of ultrasound imaging is prevalent in the diagnostic evaluation of breast cancer, with its predictive accuracy being dependent on the expertise of the specialist. Therefore, there is an urgent need to create fast and reliable ultrasound image detection algorithms to address this issue. Methods: This paper aims to compare the efficiency of six state-of-the-art, fine-tuned deep learning models that can classify breast tissue from ultrasound images into three classes: benign, malignant, and normal, using transfer learning. Additionally, the architecture of a custom model is introduced and trained from the ground up on a public dataset containing 780 images, which was further augmented to 3900 and 7800 images, respectively. What is more, the custom model is further validated on another private dataset containing 163 ultrasound images divided into two classes: benign and malignant. The pre-trained architectures used in this work are ResNet-50, Inception-V3, Inception-ResNet-V2, MobileNet-V2, VGG-16, and DenseNet-121. The performance evaluation metrics that are used in this study are as follows: Precision, Recall, F1-Score and Specificity. Results: The experimental results show that the models trained on the augmented dataset with 7800 images obtained the best performance on the test set, having 94.95 ± 0.64%, 97.69 ± 0.52%, 97.69 ± 0.13%, 97.77 ± 0.29%, 95.07 ± 0.41%, 98.11 ± 0.10%, and 96.75 ± 0.26% accuracy for the ResNet-50, MobileNet-V2, InceptionResNet-V2, VGG-16, Inception-V3, DenseNet-121, and our model, respectively. Conclusion: Our proposed model obtains competitive results, outperforming some state-of-the-art models in terms of accuracy and training time. [ABSTRACT FROM AUTHOR]

Hyperspectral images are complex images that contain more spectral dimension information than ordinary images. An increasing number of HSI classification methods are using deep learning techniques to process three-dimensional data. The Vision Transformer model is gradually occupying an important position in the field of computer vision and is being used to replace the CNN structure of the network. However, it is still in the preliminary research stage in the field of HSI. In this paper, we propose using a spectral Swin Transformer network for HSI classification, providing a new approach for the HSI field. The Swin Transformer uses group attention to enhance feature representation, and the sliding window attention calculation can take into account the contextual information of different windows, which can retain the global features of HSI and improve classification results. In our experiments, we evaluated our proposed approach on several public hyperspectral datasets and compared it with several methods. The experimental results demonstrate that our proposed model achieved test accuracies of 97.46%, 99.7%, and 99.8% on the IP, SA, and PU public HSI datasets, respectively, when using the AdamW optimizer. Our approach also shows good generalization ability when applied to new datasets. Overall, our proposed approach represents a promising direction for hyperspectral image classification using deep learning techniques. [ABSTRACT FROM AUTHOR]

Image recognition has long been one of the research hotspots in computer vision tasks. The development of deep learning is rapid in recent years, and convolutional neural networks usually need to be designed with fixed resources. If sufficient resources are available, the model can be scaled up to achieve higher accuracy, for example, VggNet, ResNet, GoogLeNet, etc. Although the accuracy of large-scale models has been improved, the following problems will occur with the expansion of model scale: (1) There may be over-fitting; (2) increasing model parameters; (3) slow model convergence. This paper proposes a design method for a modular convolutional neural network model which solves the problem of over-fitting and large model parameters by connecting multiple modules in parallel. Moreover, each module contains several submodules (three submodules in this paper) and fuses the features extracted from the submodules. The model convergence can be accelerated by using the fused features (the fused features contain more image information). In this study, we add a gate unit based on the attention mechanism to the model, which aims to optimize the structure of the model (select the optimal number of modules), allowing the model to select an optimum network structure by learning and dynamically reducing FLOPs (floating-point operations per second) of the model. Compared to VggNet, ResNet, and GoogLeNet, the structure of the model proposed in this paper is simple and the parameters are small. The proposed model achieves good results in the Kaggle datasets Cats-vs.-Dogs (99.3%), 10-Monkey Species (99.26%), and Birds-400 (99.13%). [ABSTRACT FROM AUTHOR]

Gaze estimation is an established research problem in computer vision. It has various applications in real life, from human–computer interactions to health care and virtual reality, making it more viable for the research community. Due to the significant success of deep learning techniques in other computer vision tasks—for example, image classification, object detection, object segmentation, and object tracking—deep learning-based gaze estimation has also received more attention in recent years. This paper uses a convolutional neural network (CNN) for person-specific gaze estimation. The person-specific gaze estimation utilizes a single model trained for one individual user, contrary to the commonly-used generalized models trained on multiple people's data. We utilized only low-quality images directly collected from a standard desktop webcam, so our method can be applied to any computer system equipped with such a camera without additional hardware requirements. First, we used the web camera to collect a dataset of face and eye images. Then, we tested different combinations of CNN parameters, including the learning and dropout rates. Our findings show that building a person-specific eye-tracking model produces better results with a selection of good hyperparameters when compared to universal models that are trained on multiple users' data. In particular, we achieved the best results for the left eye with 38.20 MAE (Mean Absolute Error) in pixels, the right eye with 36.01 MAE, both eyes combined with 51.18 MAE, and the whole face with 30.09 MAE, which is equivalent to approximately 1.45 degrees for the left eye, 1.37 degrees for the right eye, 1.98 degrees for both eyes combined, and 1.14 degrees for full-face images. [ABSTRACT FROM AUTHOR]

In this study, we used image recognition technology to explore different ways to improve the safety of construction workers. Three object recognition scenarios were designed for safety at a construction site, and a corresponding object recognition model was developed for each scenario. The first object recognition model checks whether there are construction workers at the site. The second object recognition model assesses the risk of falling (falling off a structure or falling down) when working at an elevated position. The third object recognition model determines whether the workers are appropriately wearing safety helmets and vests. These three models were newly created using the image data collected from the construction sites and synthetic image data collected from the virtual environment based on transfer learning. In particular, we verified an artificial intelligence model based on a virtual environment in this study. Thus, simulating and performing tests on worker falls and fall injuries, which are difficult to re-enact by humans, are efficient algorithm verification methods. The verification and synthesis data acquisition method based on a virtual environment is one of the main contributions of this study. This paper describes the overall application development approach, including the structure and method used to collect the construction site image data, structure of the training image dataset, image dataset augmentation method, and the artificial intelligence backbone model applied for transfer learning. [ABSTRACT FROM AUTHOR]

Non-Intrusive Load Monitoring (NILM) is an effective energy consumption analysis technology, which just requires voltage and current signals on the user bus. This non-invasive monitoring approach can clarify the working state of multiple loads in the building with fewer sensing devices, thus reducing the cost of energy consumption monitoring. In this paper, an NILM method combining adaptive Recurrence Plot (RP) feature extraction and deep-learning-based image recognition is proposed. Firstly, the time-series signal of current is transformed into a threshold-free RP in phase space to obtain the image features. The Euclidean norm in threshold-free RP is scaled exponentially according to the voltage and current correlation to reflect the working characteristics of different loads adaptively. Afterwards, the obtained adaptive RP features can be mapped into images using the corresponding pixel value. In the load identification stage, an advanced computer vision deep network, Hierarchical Vision Transformer using Shifted Windows (Swin-Transformer), is applied to identify the adaptive RP images. The proposed solution is extensively verified by four real, measured load signal datasets, including industrial and household power situations, covering single-phase and three-phase electrical signals. The numerical results demonstrate that the proposed NILM method based on the adaptive RP can effectively improve the accuracy of load detection. [ABSTRACT FROM AUTHOR]

Low-light images are a common phenomenon when taking photos in low-light environments with inappropriate camera equipment, leading to shortcomings such as low contrast, color distortion, uneven brightness, and high loss of detail. These shortcomings are not only subjectively annoying but also affect the performance of many computer vision systems. Enhanced low-light images can be better applied to image recognition, object detection and image segmentation. This paper proposes a novel RetinexDIP method to enhance images. Noise is considered as a factor in image decomposition using deep learning generative strategies. The involvement of noise makes the image more real, weakens the coupling relationship between the three components, avoids overfitting, and improves generalization. Extensive experiments demonstrate that our method outperforms existing methods qualitatively and quantitatively. [ABSTRACT FROM AUTHOR]

It is critical for intelligent vehicles to be capable of monitoring the health and well-being of the drivers they transport on a continuous basis. This is especially true in the case of autonomous vehicles. To address the issue, an automatic system is developed for driver's real emotion recognizer (DRER) using deep learning. The emotional values of drivers in indoor vehicles are symmetrically mapped to image design in order to investigate the characteristics of abstract expressions, expression design principles, and an experimental evaluation is conducted based on existing research on the design of driver facial expressions for intelligent products. By substituting a custom-created CNN features learning block with the base 11 layers CNN model in this paper for the development of an improved faster R-CNN face detector that detects the driver's face at a high frame per second (FPS). Transfer learning is performed in the NasNet large CNN model in order to recognize the driver's various emotions. Additionally, a custom driver emotion recognition image dataset is being developed as part of this research task. The proposed model, which is a combination of an improved faster R-CNN and transfer learning in NasNet-Large CNN architecture for DER based on facial images, enables greater accuracy than previously possible for DER based on facial images. The proposed model outperforms some recently updated state-of-the-art techniques in terms of accuracy. The proposed model achieved the following accuracy on various benchmark datasets: JAFFE 98.48%, CK+ 99.73%, FER-2013 99.95%, AffectNet 95.28%, and 99.15% on a custom-developed dataset. [ABSTRACT FROM AUTHOR]

Machine learning (ML) and especially deep learning (DL) with neural networks have demonstrated an amazing success in all sorts of AI problems, from computer vision to game playing, from natural language processing to speech and image recognition. In many ways, the approach of ML toward solving a class of problems is fundamentally different than the one followed in classical engineering, or with ontologies. While the latter rely on detailed domain knowledge and almost exhaustive search by means of static inference rules, ML adopts the view of collecting large datasets and processes this massive information through a generic learning algorithm that builds up tentative solutions. Combining the capabilities of ontology-based recommendation and ML-based techniques in a hybrid system is thus a natural and promising method to enhance semantic knowledge with statistical models. This merge could alleviate the burden of creating large, narrowly focused ontologies for complicated domains, by using probabilistic or generative models to enhance the predictions without attempting to provide a semantic support for them. In this paper, we present a novel hybrid recommendation system that blends a single architecture of classical knowledge-driven recommendations arising from a tailored ontology with recommendations generated by a data-driven approach, specifically with classifiers and a neural collaborative filtering. We show that bringing together these knowledge-driven and data-driven worlds provides some measurable improvement, enabling the transfer of semantic information to ML and, in the opposite direction, statistical knowledge to the ontology. Moreover, the novel proposed system enables the extraction of the reasoning recommendation results after updating the standard ontology with the new products and user behaviors, thus capturing the dynamic behavior of the environment of our interest. [ABSTRACT FROM AUTHOR]

Deep learning has obtained remarkable achievements in computer vision, especially image and video processing. However, in synthetic aperture radar (SAR) image recognition, the application of DNNs is usually restricted due to data insufficiency. To augment datasets, generative adversarial networks (GANs) are usually used to generate numerous photo-realistic SAR images. Although there are many pixel-level metrics to measure GAN's performance from the quality of generated SAR images, there are few measurements to evaluate whether the generated SAR images include the most representative features of the target. In this case, the classifier probably categorizes a SAR image into the corresponding class based on "wrong" criterion, i.e., "Clever Hans". In this paper, local interpretable model-agnostic explanation (LIME) is innovatively utilized to evaluate whether a generated SAR image possessed the most representative features of a specific kind of target. Firstly, LIME is used to visualize positive contributions of the input SAR image to the correct prediction of the classifier. Subsequently, these representative SAR images can be selected handily by evaluating how much the positive contribution region matches the target. Experimental results demonstrate that the proposed method can ally "Clever Hans" phenomenon greatly caused by the spurious relationship between generated SAR images and the corresponding classes. [ABSTRACT FROM AUTHOR]

As a data-driven approach, deep learning requires a large amount of annotated data for training to obtain a sufficiently accurate and generalized model, especially in the field of computer vision. However, when compared with generic object recognition datasets, aerial image datasets are more challenging to acquire and more expensive to label. Obtaining a large amount of high-quality aerial image data for object recognition and image understanding is an urgent problem. Existing studies show that synthetic data can effectively reduce the amount of training data required. Therefore, in this paper, we propose the first synthetic aerial image dataset for ship recognition, called UnityShip. This dataset contains over 100,000 synthetic images and 194,054 ship instances, including 79 different ship models in ten categories and six different large virtual scenes with different time periods, weather environments, and altitudes. The annotations include environmental information, instance-level horizontal bounding boxes, oriented bounding boxes, and the type and ID of each ship. This provides the basis for object detection, oriented object detection, fine-grained recognition, and scene recognition. To investigate the applications of UnityShip, the synthetic data were validated for model pre-training and data augmentation using three different object detection algorithms and six existing real-world ship detection datasets. Our experimental results show that for small-sized and medium-sized real-world datasets, the synthetic data achieve an improvement in model pre-training and data augmentation, showing the value and potential of synthetic data in aerial image recognition and understanding tasks. [ABSTRACT FROM AUTHOR]

Mineral recognition is of importance in geological research. Traditional mineral recognition methods need professional knowledge or special equipment, are susceptible to human experience, and are inconvenient to carry in some conditions such as in the wild. The development of computer vision provides a possibility for convenient, fast, and intelligent mineral recognition. Recently, several mineral recognition methods based on images using a neural network have been proposed for this aim. However, these methods do not exploit features extracted from the backbone network or available information of the samples in the mineral dataset sufficiently, resulting in low recognition accuracy. In this paper, a method based on feature fusion and online hard sample mining is proposed to improve recognition accuracy by using only mineral photo images. This method first fuses multi-resolution features extracted from ResNet-50 to obtain comprehensive information of mineral photos, and then proposes the weighted top-k loss to emphasize the learning of hard samples. Based on a dataset consisting of 14,986 images of 22 common minerals, the proposed method with 10-fold cross-validation achieves a Top1 accuracy of 88.01% on the validation image set, surpassing those of Inception-v3 and EfficientNet-B0 by a margin of 1.88% and 1.29%, respectively, which demonstrates the good prospect of the proposed method for convenient and reliable mineral recognition using mineral photos only. [ABSTRACT FROM AUTHOR]