Search

The performance of deep neural networks (DNNs) critically relies on high-quality annotations, while training DNNs with noisy labels remains challenging owing to their incredible capacity to memorize the entire training set. In this work, we use two synchronously trained networks to reveal that noisy labels may result in more divergent gradients when updating the parameters. To overcome this, we propose a novel co-training framework named gradient agreement learning (GAL). By dynamically evaluating the gradient agreement coefficient of every pair of parameters from two identical DNNs to determine whether to update them in the training process. GAL can effectively hinder the memorization of noisy labels. Furthermore, we utilize the pseudo labels produced by the two DNNs as the supervision for the training of another network, thereby gaining further improvement by correcting some noisy labels while overcoming the confirmation bias. Extensive experiments on various benchmark datasets demonstrate the superiority of the proposed GAL.

Pneumonia is a relatively common disease that will endanger the lives of patients if left untreated. End-to-end detection of pneumonia using neural networks will be helpful for reducing related workforce. CNN's processing of images shows remarkable performance, naturally, the use of CNN based methods for assisted reading will be a trend in modern medicine. The property of current detection algorithms is not yet satisfactory, so further research is extremely needed. In this article, we design GeminiNet to identify and localize the pneumonia in Chest X-ray (CXR) images. It uses a popular fully convolution architecture with computation shared on the entire image, combining RoI Align and PSRoI Pooling to capture global and local information and output. Our approach introduces DetNet59, a network designed specifically for detection to capture deep features. In the sixth stage of DetNet59, the structure of the retina-like convolutional layers is added to replace the fully connected layer. This structure uses the dilated convolution to extend the reconstructive field, and the convolution kernels of three different scales are used for parallel calculation to collect rich feature information. GeminiNet is validated on the RSNA dataset. We augment dataset by flipping on horizontal and vertical for the small amount of data. At IoU (Intersection over Union) =0.5, AP reached 0.4575, 0.078 higher than ResNet50, and reached 0.7758 on the AUC indicator. GeminiNet achieves 8fps in detection speed, which is better than the 7fps of the popular Faster R-CNN architecture.

A small portion of mislabeled data can easily limit the performance of deep neural networks (DNNs) due to their high capacity for memorizing random labels. Thus, robust learning from noisy labels has become a key challenge for deep learning due to inadequate datasets with high-quality annotations. Most existing methods involve training models on clean sets by dividing clean samples from noisy ones, resulting in large amounts of mislabeled data being unused. To address this problem, we propose categorizing training samples into five fine-grained clusters based on the difficulty experienced by DNN models when learning them and label correctness. A novel fine-grained confidence modeling (FGCM) framework is proposed to cluster samples into these five categories; with each cluster, FGCM decides whether to accept the cluster data as they are, accept them with label correction, or accept them as unlabeled data. By applying different strategies to the fine-grained clusters, FGCM can better exploit training data than previous methods. Extensive experiments on widely used benchmarks CIFAR-10, CIFAR-100, clothing1M, and WebVision with different ratios and types of label noise demonstrate the superiority of our FGCM.

In consideration of secure and convenient, face gains increasing attention in variety of fields during the past decades. Since human face is most accessible from our daily life and preserves the richest information, face based biometric systems are widely used in person authentication applications. However, face recognition systems are always challenged by face spoofing attacks. Although, researchers have proposed many face spoofing detection methods, which have achieved great performances, we aim to develop a method to counter face spoofing, which combines the face detection stage and face spoofing detection stage together. In this paper, we design face anti-spoofing region-based convolutional neural network (FARCNN), based on improved Faster region-based convolutional neural network (R-CNN) framework. Motivated by face detection, we regard the face spoofing detection as a three-way classification to distinguish real face, fake face and background. We extend the typical Faster R-CNN scheme by optimizing several important strategies, including roi-pooling feature fusion and adding Crystal Loss function to the original multi-task loss function. In addition, an improved Retinex based LBP is presented to handle the different illumination conditions in face spoofing detection. Finally, these two detectors are further cascaded and achieve promising performances on the benchmark databases: CASIA-FASD, REPLAY-ATTACK and OULU-NPU. Besides, for the purpose of verifying the generalization capacity of the proposed cascade detector, we perform experiments on cross-databases and the results testify the effectiveness of our proposed method.

Pedestrian detection attracts much attention from the academic community since it is an essential and significant component of autonomous driving. Despite many similarities with general object detection, pedestrian detection still has unique challenges such as background false positives and missed detections in a crowd. In this paper, we explore the value of contextual information for pedestrian detection, arguing that the contextual information is not a panacea in all cases. Despite the importance of enriching context, a method should focus on the core area where there is the target instance. To this end, we propose a pedestrian detection framework that extends proposals to enrich the context and relieves the confusions caused by contextual information through a supervised attention module. The ablation study demonstrates that the proposed extension is effective for both crowded and non-crowded instances; moreover, the attention module narrows the precision gap between these two cases. The extensive experiments on commonly used pedestrian benchmarks are conducted, which show the superior performance of our method. We achieve the average precision of 74.78% on KITTI pedestrian benchmark with the hard metric level, ranking the first places among all methods.

Deep convolutional neural networks have made significant progress in image denoising. However, in most cases, denoising methods using a single-stream structure with a single kernel size do not perform so well in integrating complementary contextual information; owing to the lack of this type of information, they may fail to reconstruct fine textures and patterns. To address this problem, we propose a multi-scale gated fusion network (MGFN) for image denoising, which learns direct end-to-end mappings from corrupted images to clean images. Our proposed network consists of several multi-scale mutually-gated (MM) blocks. In each MM block, we incorporate dilated convolution into a merge-and-run (MR) module to exploit multi-scale features in an effective way and further recognize useful features by filtration via a gating mechanism. Moreover, we propose a simple but effective loss function named dropout-loss to train the network. The extensive experiments on benchmark datasets show that our proposed method can well recover textures, yielding favorable performance against other state-of-the-art methods.

Three-dimensional (3-D) imaging sonar systems require large planar arrays, which incur hardware costs. In contrast, a cross array consisting of two perpendicular linear arrays can also support 3-D imaging while dramatically reducing the number of sensors. Moreover, the use of an aperiodic sparse array can further reduce the number of sensors efficiently. In this paper, an optimized method for sparse cross array synthesis is proposed. First, the beamforming of a cross array based on a multi-frequency algorithm is simplified for both near-field and far-field. Next, a perturbed convex optimization algorithm is proposed for sparse cross array synthesis. The method based on convex optimization utilizes a first-order Taylor expansion to create position perturbations that can optimize the beam pattern and minimize the number of active sensors. Finally, a cross array with 100 + 100 sensors is employed from which a sparse cross array with 45 + 45 sensors is obtained via the proposed method. The experimental results show that the proposed method is more effective than existing methods for obtaining optimum results for sparse cross array synthesis in both the near-field and far-field.

Pneumonia is a disease that develops rapidly and seriously threatens the survival and health of human beings. At present, the computer-aided diagnosis (CAD) of pneumonia is mostly based on binary classification algorithms that cannot provide doctors with location information. To solve this problem, this study proposes an end-to-end highly efficient algorithm for the detection of pneumonia based on a convolutional neural network—Pneumonia Yolo (PYolo). This algorithm is an improved version of the Yolov3 algorithm for X-ray image data of the lungs. Dilated convolution and an attention mechanism are used to improve the detection results of pneumonia lesions. In addition, double K-means is used to generate an anchor box to improve the localization accuracy. The algorithm obtained 46.84 mean average precision (mAP) on the X-ray image dataset provided by the Radiological Society of North America (RSNA), surpassing other detection algorithms. Thus, this study proposes an improved algorithm that can provide doctors with location information on lesions for the detection of pneumonia.