Search

We present an Augmentative Contrastive Learning for One-Shot Object Detection method that is inspired by the co-attention and co-excitation (CoAE) method. In the One-shot Object Detection (OSOD) task, the target image contains rich background information that influences the outcomes of different categories of objects more significantly. As a result, the network frequently suggests objects that do not fall under this category, leading to false detections. To learn the similarity between the target object and the query patch, we built the data augmentation module of object switch by similarity, encourage and restrain module. It primarily consists of the following three components: 1) To improve the contrast between the foreground and background, we propose an object switch module based on data augmentation and feature similarity; 2) To decrease the likelihood of false detections, we propose an encourage and restrain module that allows the network to encourage the recommendation of similar objects while restricting the recommendation of different categories of objects; and 3) To increase classification results, we use the nonlinear projector. We perform tests on the PASCAL VOC and MS COCO datasets to confirm the efficacy of our technique. [ABSTRACT FROM AUTHOR]

Cross-Domain Few-Shot Learning (CD-FSL) to recognize new categories in a new domain with few samples has attracted significant attention. Recently, task-specific CD-FSL emerges as promising research for its great generalization. However, the existing task-specific CD-FSL methods are not robust enough in task context modeling and task-specific feature learning, especially under the cross-domain setting. To tackle this problem, a Task Context Transformer and Graph Convolutional Network (TCT-GCN) method for CD-FSL is proposed. The proposed TCT-GCN constructs three modules: 1) Multi-Level Feature Fusion, which fuses domain-shared low-level features into domain-unshared high-level features; 2) Transformer-based Task Context Encoder, which models sample order-independent task context; 3) Graph Convolutional Network-based Adaptive Feature Learner, which adaptively learns task-specific features. Experiments on eight CD-FSL datasets reveal the effectiveness of our TCT-GCN. [ABSTRACT FROM AUTHOR]

Recently, a massive number of deep learning-based approaches have been successfully applied to various remote sensing image (RSI) recognition tasks. However, most existing advances of deep learning methods in the RSI field heavily rely on the features extracted by the manually designed backbone network, which severely hinders the potential of deep learning models due to the complexity of RSI and the limitation of prior knowledge. In this paper, we research a new design paradigm for the backbone architecture in RSI recognition tasks, including scene classification, land-cover classification and object detection. A novel one-shot architecture search framework based on a weight-sharing strategy and an evolutionary algorithm is proposed, called RSBNet, which consists of three stages: Firstly, a supernet constructed in a layer-wise search space is pretrained on a self-assembled large-scale RSI dataset based on an ensemble single-path training strategy. Next, the pre-trained supernet is equipped with different recognition heads through the switchable recognition module and respectively fine-tuned on the target dataset to obtain task-specific supernet. Finally, we search for the optimal backbone architecture for different recognition tasks based on the evolutionary algorithm without any network training. Extensive experiments have been conducted on five benchmark datasets for different recognition tasks, the results show the effectiveness of the proposed search paradigm and demonstrate that the searched backbone is able to flexibly adapt different RSI recognition tasks and achieve impressive performance. [ABSTRACT FROM AUTHOR]

Many traditional unsupervised feature selection algorithms utilize manifold information to mine the local structure of the data. However, the noise existing in the raw data reduces the accuracy of the manifold information of data, which affects the learning effect of the entire algorithm. In order to solve the above problems and more fully find the internal structure inside the data, this paper proposes an adaptive graph regularization and self-expression for noise-aware feature selection (ASNFS). Firstly, the algorithm adopts non-negative matrix factorization to decompose the raw data matrix, and utilizes the low-dimensional matrix generated after decomposition to replace the raw high-dimensional data matrix. This allows the algorithm to reveal some internal structural information of the raw data while reducing the dimensionality of the data. ASNFS also introduces the orthogonal basis clustering with excellent clustering effect, and the interpretability of the algorithm is enhanced. Secondly, in addition to preserving the manifold information in the low-dimensional projection subspace, the algorithm also preserves the manifold information in the non-negative matrix factorization subspace. Meanwhile, the adaptive graph regularization term added to the objective function enables the algorithm to continuously update the similarity matrix. It can effectively remove the noise inside the raw data, and prevent the occurrence of over-fitting phenomenon of the experimental results caused by the fixed similarity matrix. Finally, the similarity matrix retains the local structure information of the data with each iteration, and the results of the feature selection are reused for the construction of the similarity matrix. ASNFS adopts an alternate iterative method to optimize the objective function, which is simple and effective. Then, the algorithm complexity and convergence are analyzed. ASNFS is compared with seven feature selection algorithms on nine datasets, and the experimental results reflect the effectiveness of ASNFS in feature selection. [ABSTRACT FROM AUTHOR]

With the introduction of spatial-spectral fusion and deep learning, the classification performance of hyperspectral imagery (HSI) has been promoted greatly. For some widely used datasets, the classification accuracy almost reaches 100%. However, for hyperspectral image classification, random sampling is still the most common strategy to collect the training and test samples. Because the training and test samples are randomly selected from the same image, so they have a high correlation and the classification results are overoptimistic. Besides, random sampling is not a good choice for practical applications because we cannot always collect training and test samples from the same region. Disjoint sampling selects training and testing samples from different local regions, which will provide a more objective performance evaluation for HSI classification models. In this paper, we first show the huge classification performance difference caused by different sampling strategies with a simple experiment, then we analyze the underlying reasons from the spectral information, spatial-spectral combination, sample overlapping and spatial distance, finally, a semi-supervised feature learning method is proposed for disjoint HSI classification, in which the spatial and spectral information are exploited effectively and reasonably. The experimental results based on three HSI datasets demonstrate the effectiveness of the proposed method. [ABSTRACT FROM AUTHOR]

Most of the existing unsupervised feature selection methods learn the cluster structure through spectral clustering, and then use various regression models to introduce the data matrix into the indicator matrix to obtain feature selection matrix. In these methods, the clustering indicator matrix is usually continuous value, which is not the best choice for the matrix in terms of its supervising role in feature selection. Based on this, unsupervised feature selection via discrete spectral clustering and feature weights (FSDSC) is proposed in this paper. First, FSDSC integrates regression model and spectral clustering in a unified framework for feature selection, and introduces a feature weight matrix, which intuitively expresses the importance of each feature with its diagonal elements. Compared with the common feature selection matrix that requires constraints such as sparse regular items, the appearance of the feature weight matrix reduces the complexity of the model and simplifies the calculation process of feature evaluation. Secondly, for the value of the indicators matrix, the spectral clustering is improved to obtain a discrete clustering indicator matrix, which provides clearer guidance information for feature selection. Finally, in order to avoid trivial solutions, the transformation matrix is constrained by orthogonal constraint. The combination of the orthogonal regression model and spectral clustering enables the algorithm to perform feature selection and manifold information learning at the same time, thereby preserving the local geometric structure of data. Compared with other excellent unsupervised feature selection algorithms, the experimental results prove the effectiveness of the proposed algorithm. [ABSTRACT FROM AUTHOR]

• This paper proposes a new feature selection method called SLMEA. • The sparse transformation and pseudo label matrix are combined to optimize. • The pseudo label matrix is used to guide the learning of low-dimensional space. • The similarity matrix is constructed adaptively. • The constraint of ℓ 2 , 1 / 2 -norm is applied to the matrix. • The SLMEA can have higher clustering accuracy and normalized mutual information. Traditional feature selection algorithms usually explore the relationship between data and cluster structure in a single space, so the internal relationship obtained is not very rich, and it is not enough to select more valuable features. To solve the above problems, in order to fully mine the intrinsic correlation information in different spaces, so that the relationship between data, features and clustering structure can be explored at the same time, this paper proposes a feature selection method based on sparse and low-dimensional representation with maximum entropy adaptive graph (SLMEA). Firstly, the SLMEA combines the sparse transform representation with pseudo-label matrix learning to optimize, and uses the pseudo-label matrix to guide the learning of sparse low-dimensional space. It can not only explore the relationship between data and pseudo-labels in data space, but also mine the association between features and pseudo-labels in feature space, so as to select more discriminative features. Secondly, based on the maximum entropy theory, the similarity matrix is constructed adaptively, so that the two different manifold structures corresponding to sparse transform representation and pseudo label matrix learning can be adaptively learned and retained in the iterative process. In addition, in order to ensure the sparsity of the transformation matrix, the constraint of ℓ 2 , 1 / 2 -norm is applied to the matrix, which can better deal with the redundant features and obtain more sparse solutions. Finally, the SLMEA uses an alternate iterative update method to optimize the objective function, and carries out extensive experiments on eight mainstream datasets. Compared with seven state-of-the-art algorithms, SLMEA can have higher clustering accuracy and normalized mutual information. [ABSTRACT FROM AUTHOR]

Weakly supervised object localization (WSOL) is a challenging task that aims to localize objects in images using only image-level labels. Despite the widespread use of WSOL methods based on class activation mapping (CAM), such methods do not consider that the network may overly focus on local regions of the most interesting objects during localization, thus neglecting the overall information. To address this issue, we introduces an additional attention branch for convolutional neural networks (CNNs) that utilizes the attention mechanism of multi-layer perceptron (MLP) to enhance the network's learning of global information and supervise the feature learning of CNNs online through knowledge distillation, thereby improving the localization accuracy of WSOL. Specifically, we designs a new loss function using the generated features to combine with contrastive learning, effectively dividing the foreground and background of the image to provide more accurate pseudo-labels for subsequent classification and localization tasks. In the experiments, we tested our method on the CUB-200-2011 dataset and compared it with existing methods. The experimental results show that our method achieves good performance in WSOL tasks. [ABSTRACT FROM AUTHOR]

Pedestrian detection has emerged as a fundamental technology for autonomous cars, robotics, pedestrian search, and other applications. Although a lot of progress has been made in pedestrian detection, it is still a challenging problem in wide field-of-view surveillance videos due to the random distribution and dynamic characteristics of pedestrians. Especially in large-scale high-resolution images, there are many pedestrians and vast scale variance, so it is difficult to accurately detect all pedestrians. In order to solve this problem, sliding window is used to crop all original images to obtain pre-detection results firstly in this paper. Then, the original images are cropped again with the object as the center utilizing the label files shared in the same scene to get multi-scale images. Finally, a region NMS algorithm, a fusion strategy about the results of small images mapped into large images, is proposed to remove the redundant shredded boxes caused by cropping image. We verify the effectiveness of the proposed method with Faster R-CNN, Cascade R-CNN, IterDet and Scale-aware Fast R-CNN models on PANDA dataset. [ABSTRACT FROM AUTHOR]

Knowledge distillation (KD) builds a lightweight Student Model (SM) and trains it to approximate a large Teacher Model (TM) by exploring knowledge learned by the TM, which shows effectiveness to train lightweight CNN models. However, training a small SM to achieve better performance remains a challenging problem. Recent researches on human learning behaviors show that both the knowledge from teachers and the knowledge learning processes of teachers are significant for students. Inspired by this characteristic, in this paper, we propose a new perspective, called Progressive Mimic Learning (PML), to train lightweight CNN models by mimicking the learning trajectory of the TM. In order to obtain a more powerful SM, the useful hints in the learning process of the TM are explored. To start with, the TM learning process is divided into multiple stages, and the last state of the TM in each stage is recorded as a landmark. The learning trajectory of the TM is composed of these landmarks. Then, a landmark loss is defined to constrain the SM to progressively mimic the learning process of the TM, by employing landmarks in the learning trajectory as a training hint of the SM. Several experiments are conducted on four benchmark data sets, CIFAR-10, CIFAR-100, Fashion-MNIST, and ImageNet-10, to investigate the performance of the PML. The results show that the PML can make SMs generate more accurate predictions than SMs trained by its counterparts. [ABSTRACT FROM AUTHOR]

• A precise crop classification method of UAV hyperspectral imagery is proposed. • KTSSCC can retain the spectral and spatial information while improving the linear separability. • The KTSSOMP algorithm adaptively learns the sparse coding coefficients of KTSSCC. • There are very few parameters in our proposed KTSSCC model. Precise crop classification plays a significant role in the agriculture field. An appropriate data source for precise crop classification is high spatial resolutions hyperspectral imagery (H2 imagery) acquired by unmanned aerial vehicle (UAV). However, for imagery with many different classes of crops, crop classification of UAV H2 imagery is a huge challenge. The significant spectral diversity, spatial heterogeneity and nonlinear data structure of UAV H2 imagery results in poor spectral discriminability. To improve the discriminability, a kernel tensor slice sparse coding-based classifier (KTSSCC) is proposed for precise crop classification of UAV H2 imagery in this research. The kernel tensor representation mechanism in KTSSCC can reduce the nonlinear separation while well preserving the spectral characteristics and spatial constraints of land-covers, and thus the discriminability is greatly improved. Furthermore, this paper puts forward the kernel tensor slice sparse orthogonal matching pursuit (KTSSOMP) algorithm to optimize kernel tensor slice sparse coding in the spectral space, which greatly reduces the computation cost. Moreover, there are very few parameters to be tuned in our proposed model. We assess the performance of KTSSCC on two real UAV hyperspectral imagery datasets, and find that, based on visual and quantitative results, it provides satisfactory crop classification results and outperforms the state-of-the-art approaches. [ABSTRACT FROM AUTHOR]