In this paper, we present a novel discriminative dictionary learning (DDL) method for image classification. The local structural relationship between samples is first built by the Laplacian eigenmaps (LE), and then integrated into the basic DDL frame to suppress inter-class ambiguity in the feature space. Moreover, in order to improve the discriminative ability of the dictionary, the category label information of training samples is formulated into the objective function of dictionary learning by considering the discriminative promotion term. Thus, the data points of original samples are transformed into a new feature space, in which the points from different categories are expected to be far apart. The test results based on the real dataset indicate the effectiveness of this method.

Many deep convolutional neural network (CNN) inference accelerators on the field-programmable gate array (FPGA) platform have been widely adopted due to their low power consumption and high performance. In this paper, we develop the following to improve performance and power efficiency. First, we use a high bandwidth memory (HBM) to expand the bandwidth of data transmission between the off-chip memory and the accelerator. Second, a fully-pipelined manner, which consists of pipelined inter-layer computation and a pipelined computation engine, is implemented to decrease idle time among layers. Third, a multi-core architecture with shared-dual buffers is designed to reduce off-chip memory access and maximize the throughput. We designed the proposed accelerator on the Xilinx Alveo U280 platform with in-depth Verilog HDL instead of high-level synthesis as the previous works and explored the VGG-16 model to verify the system during our experiment. With a similar accelerator architecture, the experimental results demonstrate that the memory bandwidth of HBM is 13.2× better than DDR4. Compared with other accelerators in terms of throughput, our accelerator is 1.9×/1.65×/11.9× better than FPGA+HBM2 based/low batch size (4) GPGPU/low batch size (4) CPU. Compared with the previous DDR+FPGA/DDR+GPGPU/DDR+CPU based accelerators in terms of power efficiency, our proposed system provides 1.4-1.7×/1.7-12.6×/6.6-37.1× improvement with the large-scale CNN model.

In various studies of attacks on autonomous vehicles (AVs), a phantom attack in which advanced driver assistance system (ADAS) misclassifies a fake object created by an adversary as a real object has been proposed. In this paper, we propose F-GhostBusters, which is an improved version of GhostBusters that detects phantom attacks. The proposed model uses a new feature, i.e, frequency of images. Experimental results show that F-GhostBusters not only improves the detection performance of GhostBusters but also can complement the accuracy against adversarial examples.

In recent years, deep neural networks (DNNs) have made a significant impact on a variety of research fields and applications. One drawback of DNNs is that it requires a huge amount of dataset for training. Since it is very expensive to ask experts to label the data, many non-expert data collection methods such as web crawling have been proposed. However, dataset created by non-experts often contain corrupted labels, and DNNs trained on such dataset are unreliable. Since DNNs have an enormous number of parameters, it tends to overfit to noisy labels, resulting in poor generalization performance. This problem is called Learning with Noisy labels (LNL). Recent studies showed that DNNs are robust to the noisy labels in the early stage of learning before over-fitting to noisy labels because DNNs learn the simple patterns first. Therefore DNNs tend to output true labels for samples with noisy labels in the early stage of learning, and the number of false predictions for samples with noisy labels is higher than for samples with clean labels. Based on these observations, we propose a new sample selection approach for LNL using the number of false predictions. Our method periodically collects the records of false predictions during training, and select samples with a low number of false predictions from the recent records. Then our method iteratively performs sample selection and training a DNNs model using the updated dataset. Since the model is trained with more clean samples and records more accurate false predictions for sample selection, the generalization performance of the model gradually increases. We evaluated our method on two benchmark datasets, CIFAR-10 and CIFAR-100 with synthetically generated noisy labels, and the obtained results which are better than or comparative to the-state-of-the-art approaches.

We introduce a hybrid Bayesian-convolutional neural network (hyBCNN) for improving the robustness against adversarial attacks and decreasing the computation time in the Bayesian inference phase. Our hyBCNN models are built from a part of BNN and CNN. Based on pre-trained CNNs, we only replace convolutional layers and activation function of the initial stage of CNNs with our Bayesian convolutional (BC) and Bayesian activation (BA) layers as a term of transfer learning. We keep the remainder of CNNs unchanged. We adopt the Bayes without Bayesian Learning (BwoBL) algorithm for hyBCNN networks to execute Bayesian inference towards adversarial robustness. Our proposal outperforms adversarial training and robust activation function, which are currently the outstanding defense methods of CNNs in the resistance to adversarial attacks such as PGD and C&W. Moreover, the proposed architecture with BwoBL can easily integrate into any pre-trained CNN, especially in scaling networks, e.g., ResNet and EfficientNet, with better performance on large-scale datasets. In particular, under l∞ norm PGD attack of pixel perturbation ε=4/255 with 100 iterations on ImageNet, our best hyBCNN EfficientNet reaches 93.92% top-5 accuracy without additional training.

A triple loss based framework for generalized zero-shot learning is presented in this letter. The approach learns a shared latent space for image features and attributes by using aligned variational autoencoders and variants of triplet loss. Then we train a classifier in the latent space. The experimental results demonstrate that the proposed framework achieves great improvement.

Virtual Adversarial Training (VAT) has shown impressive results among recently developed regularization methods called consistency regularization. VAT utilizes adversarial samples, generated by injecting perturbation in the input space, for training and thereby enhances the generalization ability of a classifier. However, such adversarial samples can be generated only within a very small area around the input data point, which limits the adversarial effectiveness of such samples. To address this problem we propose LVAT (Latent space VAT), which injects perturbation in the latent space instead of the input space. LVAT can generate adversarial samples flexibly, resulting in more adverse effect and thus more effective regularization. The latent space is built by a generative model, and in this paper we examine two different type of models: variational auto-encoder and normalizing flow, specifically Glow. We evaluated the performance of our method in both supervised and semi-supervised learning scenarios for an image classification task using SVHN and CIFAR-10 datasets. In our evaluation, we found that our method outperforms VAT and other state-of-the-art methods.

Fine-grained image classification is a difficult problem, and previous studies mainly overcome this problem by locating multiple discriminative regions in different scales and then aggregating complementary information explored from the located regions. However, locating discriminative regions introduces heavy overhead and is not suitable for real-world application. In this paper, we propose the recursive multi-scale channel-spatial attention module (RMCSAM) for addressing this problem. Following the experience of previous research on fine-grained image classification, RMCSAM explores multi-scale attentional information. However, the attentional information is explored by recursively refining the deep feature maps of a convolutional neural network (CNN) to better correspond to multi-scale channel-wise and spatial-wise attention, instead of localizing attention regions. In this way, RMCSAM provides a lightweight module that can be inserted into standard CNNs. Experimental results show that RMCSAM can improve the classification accuracy and attention capturing ability over baselines. Also, RMCSAM performs better than other state-of-the-art attention modules in fine-grained image classification, and is complementary to some state-of-the-art approaches for fine-grained image classification. Code is available at https://github.com/Dichao-Liu/Recursive-Multi-Scale-Channel-Spatial-Attention-Module.

In the recent years, deep learning has achieved significant results in various areas of machine learning. Deep learning requires a huge amount of data to train a model, and data collection techniques such as web crawling have been developed. However, there is a risk that these data collection techniques may generate incorrect labels. If a deep learning model for image classification is trained on a dataset with noisy labels, the generalization performance significantly decreases. This problem is called Learning with Noisy Labels (LNL). One of the recent researches on LNL, called DivideMix [1], has successfully divided the dataset into samples with clean labels and ones with noisy labels by modeling loss distribution of all training samples with a two-component Mixture Gaussian model (GMM). Then it treats the divided dataset as labeled and unlabeled samples and trains the classification model in a semi-supervised manner. Since the selected samples have lower loss values and are easy to classify, training models are in a risk of overfitting to the simple pattern during training. To train the classification model without overfitting to the simple patterns, we propose to introduce consistency regularization on the selected samples by GMM. The consistency regularization perturbs input images and encourages model to outputs the same value to the perturbed images and the original images. The classification model simultaneously receives the samples selected as clean and their perturbed ones, and it achieves higher generalization performance with less overfitting to the selected samples. We evaluated our method with synthetically generated noisy labels on CIFAR-10 and CIFAR-100 and obtained results that are comparable or better than the state-of-the-art method.

Local discriminative regions play important roles in fine-grained image analysis tasks. How to locate local discriminative regions with only category label and learn discriminative representation from these regions have been hot spots. In our work, we propose Searching Discriminative Regions (SDR) and Learning Discriminative Regions (LDR) method to search and learn local discriminative regions in images. The SDR method adopts attention mechanism to iteratively search for high-response regions in images, and uses this as a clue to locate local discriminative regions. Moreover, the LDR method is proposed to learn compact within category and sparse between categories representation from the raw image and local images. Experimental results show that our proposed approach achieves excellent performance in both fine-grained image retrieval and classification tasks, which demonstrates its effectiveness.

Adversarial attacks are viewed as a danger to Deep Neural Networks (DNNs), which reveal a weakness of deep learning models in security-critical applications. Recent findings have been presented adversarial training as an outstanding defense method against adversaries. Nonetheless, adversarial training is a challenge with respect to big datasets and large networks. It is believed that, unless making DNN architectures larger, DNNs would be hard to strengthen the robustness to adversarial examples. In order to avoid iteratively adversarial training, our algorithm is Bayes without Bayesian Learning (BwoBL) that performs the ensemble inference to improve the robustness. As an application of transfer learning, we use learned parameters of pretrained DNNs to build Bayesian Neural Networks (BNNs) and focus on Bayesian inference without costing Bayesian learning. In comparison with no adversarial training, our method is more robust than activation functions designed to enhance adversarial robustness. Moreover, BwoBL can easily integrate into any pretrained DNN, not only Convolutional Neural Networks (CNNs) but also other DNNs, such as Self-Attention Networks (SANs) that outperform convolutional counterparts. BwoBL is also convenient to apply to scaling networks, e.g., ResNet and EfficientNet, with better performance. Especially, our algorithm employs a variety of DNN architectures to construct BNNs against a diversity of adversarial attacks on a large-scale dataset. In particular, under l∞ norm PGD attack of pixel perturbation ε=4/255 with 100 iterations on ImageNet, our proposal in ResNets, SANs, and EfficientNets increase by 58.18% top-5 accuracy on average, which are combined with naturally pretrained ResNets, SANs, and EfficientNets. This enhancement is 62.26% on average below l2 norm C&W attack. The combination of our proposed method with pretrained EfficientNets on both natural and adversarial images (EfficientNet-ADV) drastically boosts the robustness resisting PGD and C&W attacks without additional training. Our EfficientNet-ADV-B7 achieves the cutting-edge top-5 accuracy, which is 92.14% and 94.20% on adversarial ImageNet generated by powerful PGD and C&W attacks, respectively.

Differentiable neural architecture search (DARTS) is now a widely disseminated weight-sharing neural architecture search method and it consists of two stages: search and evaluation. However, the original DARTS suffers from some well-known shortcomings. Firstly, the width and depth of the network, as well as the operation of two stages are discontinuous, which causes a performance collapse. Secondly, DARTS has a high computational overhead. In this paper, we propose a synchronous progressive approach to solve the discontinuity problem for network depth and width and we use the 0-1 loss function to alleviate the discontinuity problem caused by the discretization of operation. The computational overhead is reduced by using the partial channel connection. Besides, we also discuss and propose a solution to the aggregation of skip operations during the search process of DARTS. We conduct extensive experiments on CIFAR-10 and WANFANG datasets, specifically, our approach reduces search time significantly (from 1.5 to 0.1 GPU days) and improves the accuracy of image recognition.

Gait recognition distinguishes one individual from others according to the natural patterns of human gaits. Gait recognition is a challenging signal processing technology for biometric identification due to the ambiguity of contours and the complex feature extraction procedure. In this work, we proposed a new model - the convolutional neural network (CNN) joint attention mechanism (CJAM) - to classify the gait sequences and conduct person identification using the CASIA-A and CASIA-B gait datasets. The CNN model has the ability to extract gait features, and the attention mechanism continuously focuses on the most discriminative area to achieve person identification. We present a comprehensive transformation from gait image preprocessing to final identification. The results from 12 experiments show that the new attention model leads to a lower error rate than others. The CJAM model improved the 3D-CNN, CNN-LSTM (long short-term memory), and the simple CNN by 8.44%, 2.94% and 1.45%, respectively.

Convolutional neural networks (CNNs) have dominated a range of applications, from advanced manufacturing to autonomous cars. For energy cost-efficiency, developing low-power hardware for CNNs is a research trend. Due to the large input size, the first few convolutional layers generally consume most latency and hardware resources on hardware design. To address these challenges, this paper proposes an innovative architecture named SLIT to extract feature maps and reconstruct the first few layers on CNNs. In this reconstruction approach, total multiply-accumulate operations are eliminated on the first layers. We evaluate new topology with MNIST, CIFAR, SVHN, and ImageNet datasets on image classification application. Latency and hardware resources of the inference step are evaluated on the chip ZC7Z020-1CLG484C FPGA with Lenet-5 and VGG schemes. On the Lenet-5 scheme, our architecture reduces 39% of latency and 70% of hardware resources with a 0.456 W power consumption compared to previous works. Even though the VGG models perform with a 10% reduction in hardware resources and latency, we hope our overall results will potentially give a new impetus for future studies to reach a higher optimization on hardware design. Notably, the SLIT architecture efficiently merges with most popular CNNs at a slightly sacrificing accuracy of a factor of 0.27% on MNIST, ranging from 0.5% to 1.5% on CIFAR, approximately 2.2% on ImageNet, and remaining the same on SVHN databases.

Automatic tracking and classification are essential for studying the behaviors of wild animals. Owing to dynamic far-shooting photos, the occlusion problem, protective coloration, the background noise is irregular interference for designing a computerized algorithm for reducing human labeling resources. Moreover, wild dolphin images are hard-acquired by on-the-spot investigations, which takes a lot of waiting time and hardly sets the fixed camera to automatic monitoring dolphins on the ocean in several days. It is challenging tasks to detect well and classify a dolphin from polluted photos by a single famous deep learning method in a small dataset. Therefore, in this study, we propose a generic Cascade Small Object Detection (CSOD) algorithm for dolphin detection to handle small object problems and develop visualization to backbone based classification (V2BC) for removing noise, highlighting features of dolphin and classifying the name of dolphin. The architecture of CSOD consists of the P-net and the F-net. The P-net uses the crude Yolov3 detector to be a core network to predict all the regions of interest (ROIs) at lower resolution images. Then, the F-net, which is more robust, is applied to capture the ROIs from high-resolution photos to solve single detector problems. Moreover, a visualization to backbone based classification (V2BC) method focuses on extracting significant regions of occluded dolphin and design significant post-processing by referencing the backbone of dolphins to facilitate for classification. Compared to the state of the art methods, including faster-rcnn, yolov3 detection and Alexnet, the Vgg, and the Resnet classification. All experiments show that the proposed algorithm based on CSOD and V2BC has an excellent performance in dolphin detection and classification. Consequently, compared to the related works of classification, the accuracy of the proposed designation is over 14% higher. Moreover, our proposed CSOD detection system has 42% higher performance than that of the original Yolov3 architecture.

A multi-task convolutional neural network leading to high performance and interpretability via attribute estimation is presented in this letter. Our method can provide interpretation of the classification results of CNNs by outputting attributes that explain elements of objects as a judgement reason of CNNs in the middle layer. Furthermore, the proposed network uses the estimated attributes for the following prediction of classes. Consequently, construction of a novel multi-task CNN with improvements in both of the interpretability and classification performance is realized.

This Letter proposes a autoencoder model supervised by semantic similarity for zero-shot learning. With the help of semantic similarity vectors of seen and unseen classes and the classification branch, our experimental results on two datasets are 7.3% and 4% better than the state-of-the-art on conventional zero-shot learning in terms of the averaged top-1 accuracy.

Due to the superior performance, deep learning has been widely applied to various applications, including image classification, bioinformatics, and cybersecurity. Nevertheless, the research investigations on deep learning in the adversarial environment are still on their preliminary stage. The emerging adversarial learning methods, e.g., generative adversarial networks, have introduced two vital questions: to what degree the security of deep learning with the presence of adversarial examples is; how to evaluate the performance of deep learning models in adversarial environment, thus, to raise security advice such that the selected application system based on deep learning is resistant to adversarial examples. To see the answers, we leverage image classification as an example application scenario to propose a framework of Evaluating Deep Learning for Image Classification (EDLIC) to conduct comprehensively quantitative analysis. Moreover, we introduce a set of evaluating metrics to measure the performance of different attacking and defensive techniques. After that, we conduct extensive experiments towards the performance of deep learning for image classification under different adversarial environments to validate the scalability of EDLIC. Finally, we give some advice about the selection of deep learning models for image classification based on these comparative results.

Synthetic aperture radar (SAR) image classification is a popular yet challenging research topic in the field of SAR image interpretation. This paper presents a new classification method based on extreme learning machine (ELM) and the superpixel-guided composite kernels (SGCK). By introducing the generalized likelihood ratio (GLR) similarity, a modified simple linear iterative clustering (SLIC) algorithm is firstly developed to generate superpixel for SAR image. Instead of using a fixed-size region, the shape-adaptive superpixel is used to exploit the spatial information, which is effective to classify the pixels in the detailed and near-edge regions. Following the framework of composite kernels, the SGCK is constructed base on the spatial information and backscatter intensity information. Finally, the SGCK is incorporated an ELM classifier. Experimental results on both simulated SAR image and real SAR image demonstrate that the proposed framework is superior to some traditional classification methods.

Two different types of representations, such as an image and its manually-assigned corresponding labels, generally have complex and strong relationships to each other. In this paper, we represent such deep relationships between two different types of visible variables using an energy-based probabilistic model, called a deep relational model (DRM) to improve the prediction accuracies. A DRM stacks several layers from one visible layer on to another visible layer, sandwiching several hidden layers between them. As with restricted Boltzmann machines (RBMs) and deep Boltzmann machines (DBMs), all connections (weights) between two adjacent layers are undirected. During maximum likelihood (ML) -based training, the network attempts to capture the latent complex relationships between two visible variables with its deep architecture. Unlike deep neural networks (DNNs), 1) the DRM is a totally generative model and 2) allows us to generate one visible variables given the other, and 2) the parameters can be optimized in a probabilistic manner. The DRM can be also fine-tuned using DNNs, like deep belief nets (DBNs) or DBMs pre-training. This paper presents experiments conduced to evaluate the performance of a DRM in image recognition and generation tasks using the MNIST data set. In the image recognition experiments, we observed that the DRM outperformed DNNs even without fine-tuning. In the image generation experiments, we obtained much more realistic images generated from the DRM more than those from the other generative models.