Candidate Boxes Research Articles

Underwater object tracker: UOSTrack for marine organism grasping of underwater vehicles

A visual single-object tracker is an indispensable component of underwater vehicles (UVs) in marine organism grasping tasks. Its accuracy and stability are imperative to guide the UVs to perform grasping behavior. Although single-object trackers show competitive performance in the challenge of underwater image degradation, there are still issues with sample imbalance and exclusion of similar objects that need to be addressed for application in marine organism grasping. This paper proposes Underwater OSTrack (UOSTrack), which consists of underwater image and open-air sequence hybrid training (UOHT), and motion-based post-processing (MBPP). The UOHT training paradigm is designed to train the sample-imbalanced underwater tracker so that the tracker is exposed to a great number of underwater domain training samples and learns the feature expressions. The MBPP paradigm is proposed to exclude similar objects. It uses the estimation box predicted with a Kalman filter and the candidate boxes in the response map to relocate the lost tracked object in the candidate area. UOSTrack achieves an average performance improvement of 4.41% and 7.98% maximum compared to state-of-the-art methods on various benchmarks, respectively. Field experiments have verified the accuracy and stability of our proposed UOSTrack for UVs in marine organism grasping tasks. More details can be found at https://github.com/LiYunfengLYF/UOSTrack.

De-biased Teacher: Rethinking IoU Matching for Semi-supervised Object Detection

Most of the recent research in semi-supervised object detection follows the pseudo-labeling paradigm evolved from the semi-supervised image classification task. However, the training paradigm of the two-stage object detector inevitably makes the pseudo-label learning process for unlabeled images full of bias. Specifically, the IoU matching scheme used for selecting and labeling candidate boxes is based on the assumption that the matching source~(ground truth) is accurate enough in terms of the number of objects, object position and object category. Obviously, pseudo-labels generated for unlabeled images cannot satisfy such a strong assumption, which makes the produced training proposals extremely unreliable and thus severely spoil the follow-up training. To de-bias the training proposals generated by the pseudo-label-based IoU matching, we propose a general framework -- De-biased Teacher, which abandons both the IoU matching and pseudo labeling processes by directly generating favorable training proposals for consistency regularization between the weak/strong augmented image pairs. Moreover, a distribution-based refinement scheme is designed to eliminate the scattered class predictions of significantly low values for higher efficiency. Extensive experiments demonstrate that the proposed De-biased Teacher consistently outperforms other state-of-the-art methods on the MS-COCO and PASCAL VOC benchmarks. Source codes are available at https://github.com/wkfdb/De-biased-Teracher.

Non-Destructive Internal Defect Detection of In-Shell Walnuts by X-ray Technology Based on Improved Faster R-CNN

The purpose of this study was to achieve non-destructive detection of the internal defects of in-shell walnuts using X-ray radiography technology based on improved Faster R-CNN network model. First, the FPN structure was added to the feature-extraction layer to extract richer image information. Then, ROI Align was used instead of ROI Pooling for eliminating the localization bias problem caused by the quantization operation. Finally, the Softer-NMS module was introduced to the final regression layer with the predicted bounding box for improving the localization accuracy of the candidate boxes. The results of the study indicated that the proposed network model can effectively identify internal defects of in-shell walnuts. Specifically, the discrimination accuracies of the in-shell sound, shriveled, and empty-shell walnuts were 96.14%, 91.72%, and 94.80%, respectively, and the highest overall accuracy was 94.22%. Compared to the original Faster R-CNN network model, the improved Faster R-CNN model achieved an increase of 5.86% in mAP and 5.65% in F1-value. Consequently, the proposed method can be applied for the in-shell walnuts with shriveled and empty-shell defects.

A Multi-Stage Faster RCNN-Based iSPLInception for Skin Disease Classification Using Novel Optimization.

Nowadays, skin cancer is considered a serious disorder in which early identification and treatment of the disease are essential to ensure the stability of the patients. Several existing skin cancer detection methods are introduced by employing deep learning (DL) to perform skin disease classification. Convolutional neural networks (CNNs) can classify melanoma skin cancer images. But, it suffers from an overfitting problem. Therefore, to overcome this problem and to classify both benign and malignant tumors efficiently, the multi-stage faster RCNN-based iSPLInception (MFRCNN-iSPLI) method is proposed. Then, the test dataset is used for evaluating the proposed model performance. The faster RCNN is employed directly to perform image classification. This may heavily raise computation time and network complications. So, the iSPLInception model is applied in the multi-stage classification. In this, the iSPLInception model is formulated using the Inception-ResNet design. For candidate box deletion, the prairie dog optimization algorithm is utilized. We have utilized two skin disease datasets, namely, ISIC 2019 Skin lesion image classification and the HAM10000 dataset for conducting experimental results. The methods' accuracy, precision, recall, and F1 score values are calculated, and the results are compared with the existing methods such as CNN, hybrid DL, Inception v3, and VGG19. With 95.82% accuracy, 96.85% precision, 96.52% recall, and 0.95% F1 score values, the output analysis of each measure verified the prediction and classification effectiveness of the method.

An Area-Efficient Accelerator for Non-Maximum Suppression

Non-maximum suppression (NMS) is an essential part of the post-processing of object detectors. Most object detection models require NMS algorithms to filter overlapping candidate boxes belonging to the same object and reserve the box that best represents the object as the object representative box. However, the hardware for the standard NMS algorithm has some disadvantages, such as high complexity (especially when box numbers are significant), high latency, large area, and high power consumption. To solve these problems, we propose an efficient parallel hardware architecture, which uses a new sorting circuit with a ping-pong buffer and a new retention mechanism of the candidate box for the new NMS algorithm called distance over side-NMS (DoS-NMS). This architecture uses a PE group with a voting mechanism to simplify the algorithm for reducing the latency and area. Additionally, the PE group computes the DoS (non-intersection over union) of the candidate box and multiple object representatives in parallel, significantly reducing algorithm complexity and memory access cost. Experiments indicated that the algorithm runs on the chip with area of 0.75mm2, power consumption of 68.41 mW, and normalized area efficiency that is 3.72 and 4.84 times higher than the two state-of-the-art methods, respectively.

An unsupervised aerial tracking method of camouflaged targets in complex environments

With the widespread use of UAVs and the development of various technologies related to computer vision, much attention has been devoted to UAV-based target tracking. For the task of tracking camouflaged people in complex environments, there is a shortage of annotated datasets and poor backbone feature extraction network capabilities, which will cause tracking drift and tracking lag. To achieve unsupervised UAV target tracking under unsupervised conditions, a new target tracking model has been designed. First, we roughly detect moving targets using a camouflaged segmentation network, and then apply dynamic programming to generate smooth candidate boxes. Secondly, a deeper convolutional neural network is designed for feature extraction to replace the original backbone network, and the residual structure is modified to ensure tracking effectiveness. Extensive comparative experiments have shown that the proposed unsupervised method performs the task of UAV target tracking well.

Moving Object Detection for Complex Scenes by Merging BG Modeling and Deep Learning Method

Abstract In recent years, many studies have attempted to use deep learning for moving object detection. Some research also combines object detection methods with traditional background modeling. However, this approach may run into some problems with parameter settings and weight imbalances. In order to solve the aforementioned problems, this paper proposes a new way to combine ViBe and Faster-RCNN for moving object detection. To be more specific, our approach is to confine the candidate boxes to only retain the area containing moving objects through traditional background modeling. Furthermore, in order to make the detection able to more accurately filter out the static object, the probability of each region proposal then being retained. In this paper, we compare four famous methods, namely GMM and ViBe for the traditional methods, and DeepBS and SFEN for the deep learning-based methods. The result of the experiment shows that the proposed method has the best overall performance score among all methods. The proposed method is also robust to the dynamic background and environmental changes and is able to separate stationary objects from moving objects. Especially the overall F-measure with the CDNET 2014 dataset (like in the dynamic background and intermittent object motion cases) was 0,8572.

Research on a Traffic Sign Recognition Method under Small Sample Conditions.

Traffic signs are updated quickly, and there image acquisition and labeling work requires a lot of manpower and material resources, so it is difficult to provide a large number of training samples for high-precision recognition. Aiming at this problem, a traffic sign recognition method based on FSOD (few-shot object learning) is proposed. This method adjusts the backbone network of the original model and introduces dropout, which improves the detection accuracy and reduces the risk of overfitting. Secondly, an RPN (region proposal network) with improved attention mechanism is proposed to generate more accurate target candidate boxes by selectively enhancing some features. Finally, the FPN (feature pyramid network) is introduced for multi-scale feature extraction, and the feature map with higher semantic information but lower resolution is merged with the feature map with higher resolution but weaker semantic information, which further improves the detection accuracy. Compared with the baseline model, the improved algorithm improves the 5-way 3-shot and 5-way 5-shot tasks by 4.27% and 1.64%, respectively. We apply the model structure to the PASCAL VOC dataset. The results show that this method is superior to some current few-shot object detection algorithms.

A novel head network and group normalisation help track more accurately

AbstractMost existing Siamese Network trackers rely on a predefined anchor box to predict object position. However, they require complicated hyperparameter settings. The authors directly forecast the object boundary by using the fully convolutional network as the head of the tracking network to solve this issue and simplify the application. The end‐to‐end design avoids the setting of hyperparameters and candidate boxes. The authors also discovered that the validation loss decreased less than the training loss throughout Siamese training. The authors changed the normalisation layer from batch normalisation to group normalisation to solve this issue. It solved the problem that the loss function is difficult to decrease and increased training efficiency. Test experiments on the tracking dataset, including VOT2019 and GOT10k, show that the authors’ network outperforms DaSiamRPN and SiamFC regarding precision under the same network size and runs at 24 FPS on an AMD 4800 CPU. It also runs at 306 FPS on a 3090 GPU.

Auditory Receptive Field Net Based Automatic Snore Detection for Wearable Devices.

Although obstructive sleep apnea and hypopnea syndrome (OSAHS) is a common sleep disease, it is sometimes difficult to be detected in time because of the inconvenience of polysomnography (PSG) examination. Since snoring is one of the earliest symptoms of OSAHS, it can be used for early OSAHS prediction. With the recent development of wearable and IoT sensors, we proposed a deep learning-based accurate snore detection model for long-term home monitoring of snoring during sleep. To enhance the discriminability of features between snoring and non-snoring events, an auditory receptive field (ARF) net was proposed and integrated into the feature extraction network. Based on the feature maps derived by the feature extraction network, the detection model predicted a series of candidate boxes and corresponding confidence scores for each candidate box, which denoted whether the candidate box contained a snore event from the input sound waveforms. A snore detection dataset with a total duration of more than 4600 min was developed to evaluate the proposed model. The experimental results on this dataset revealed that the proposed model outperformed other traditional approaches and deep learning models.

A fusion-based approach to deep-learning and edge-cutting algorithms for identification and color recognition of traffic lights

Abstract The detection and color recognition of traffic lights should be the foundation for the capture of illegal driving practices. However, it may be difficult to recognize lights of different colors in intricate and unpredictable surroundings. This study implements a traffic light detection and recognition scheme that can be used for intelligent traffic. First, the images obtained from the speed camera should be pre-segmented. Then the traffic lights with colors are detected by the YOLOv5 model trained based on the image-enhancement dataset. Next, the candidate boxes of traffic lights are edge detected and clipped out of multiple lamp panels in missing video frames. Finally, the color of the candidate boxes will be determined by the lamp panel with the greatest number of bright pixels. This finding shows that the fusion-based approach performs better than a single-based algorithm for identification and color recognition of traffic lights under varying illumination and weather circumstances.

A Dense-Aware Cross-splitNet for Object Detection and Recognition

Object detection and recognition is widely used in various fields and have become key technologies in computer vision. The distribution of objects in natural images can be roughly divided into densely stacked objects and scattered objects. Due to the incomplete attributes or features of some objects in densely stacked distributions, some object detectors have missed local area details or low detection accuracy. In this paper, we propose Cross-splitNet, a novel cross-split method for dense object detection and recognition based on candidate box generation. First, an adaptive feature extraction network is constructed. Different datasets are input into convolutional neural networks with various depths, the generalization of the model. Then, the proposed cross-split algorithm is introduced to guide the different deep networks to learn features of images with various densities, according to intermediate object density classification results. Finally, we adopt a feature pyramid network (FPN) subnet to perform multi-scale feature extraction while retaining lower-layer object information and physical characteristics. The model was trained on the COCO 17, VOC 12, and VOC 07 datasets, which contain a large number of object categories. Our network was compared with several two-stage detectors, and the results show that our model achieved an average precision ( <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">AP</i> ) of 0.819 at 22.9 frames per second ( <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">FPS</i> ) on the VOC 07+12 dataset. The mean average precision ( <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">mAP</i> ) of the object detection model with R50+R2-101 backbones on the COCO dataset was increased by 1.9%.

High-precision real-time UAV target recognition based on improved YOLOv4

In recent years, unmanned aerial vehicles (UAVs) have gained widespread use in both military and civilian fields with the advancement of aviation technology and improved communication capabilities. However, the phenomenon of unauthorized UAV flights, or “black flying”, poses a serious threat to the safe flight of aircraft in airspace and public safety. To effectively interfere with and attack UAV targets, it is crucial to enhance the detection and identification of “low, slow and small” UAVs. This study focuses on achieving high-precision and lightweight detection and identification of four-rotor, six-rotor, and fixed-wing UAVs in low-altitude complex environments. By combining deep learning target detection with superresolution feature enhancement, a lightweight UAV detection model is designed and field-tested for verification. To address the challenge of detecting small UAV targets with limited information, the feature fusion network is enhanced based on the traditional YOLOv4 algorithm to improve the detection ability of small targets via small target enhancement and candidate box adjustment. The feasibility of the improved network is quantitatively and qualitatively analyzed. Channel pruning and layer pruning are then applied to the network, significantly reducing its depth and width and realizing a lightweight network. Finally, reasoning quantification is conducted on the embedded platform to enable end-side deployment of the target detection algorithm.

LiDAR-only 3D object detection based on spatial context

LiDAR-based 3D Object detection is one of the popular topics in recent years, and it is widely used in the fields of autonomous driving and robot controlling. However, due to the scanning pattern of LiDAR, the point clouds of objects at far distance are sparse and more difficult to be detected. To solve this problem, we propose a two-stage network based on spatial context information, named SC-RCNN (Spatial Context RCNN), for object detection in 3D point cloud scenes. SC-RCNN first uses a backbone with sparse convolutions and submanifold sparse convolutions to extract the voxel features of point scenes and generate a series of candidate boxes. For the sparsity of far-distance point clouds, we design the local grid point pooling (LGP Pooling) to extract features and spatial context information around candidate regions for subsequent box refinement. In addition, we propose the pyramid candidate box augmentation (PCB Augmentation) to expand the candidate boxes with a multi-scale style, enriching the feature encoding. The experimental results show that SC-RCNN significantly outperforms previous methods on KITTI dataset and Waymo dataset, and is particularly robust to the sparsity of point clouds.

GPNet: Key Point Generation Auxiliary Network for Object Detection

AbstractFor existing object detectors, anchor‐based detectors lack global information, while anchor‐free detectors based on key points lack prior position information. The above issues may lead to the imbalance of detection accuracy and shape robustness. In order to alleviate the above contradictions, a new auxiliary network key point generation network (GPNet) is proposed to improve the performance of existing object detectors. Specifically, a series of key points are generated through ground truth (GT) supervision to obtain more global information. These key points generate pseudo boxes (Pbox) with learnable parameters. Pbox has more specific prior information than the manually designed candidate boxes. The scale information of the Pbox is embed into the classification branch to obtain a more appropriate receptive field. In addition, a novel improved strategy for label assignment by combining Pbox and GT to enhance the ability to classify positive and negative samples is proposed. Extensive experiments on multiple dense prediction methods validate the effectiveness of GPNet, with a performance improvement of 1.5 AP over baseline. In particular, with ResNext‐101‐64× 4d‐DCN as the backbone, this method achieves 49.5 AP with single‐scale testing.

Power Operation Violation Identification Method Based on Point Cloud Data Preprocessing and Deep Learning under the Architecture of IoT

Aiming at the problems of low recognition accuracy and large memory occupation when using point cloud information for power operation violation, A power operation violation recognition method based on point cloud data preprocessing and deep learning under the architecture of Internet of things (IoT) is proposed. First, voxel filtering and statistical filtering methods are used to properly simplify the power operation point cloud data on the premise of ensuring the quality of reverse modeling, and the moving least square method is used to smooth the point cloud to obtain a complete and closed three-dimensional model; second, the process of power operation violation behavior recognition is divided into two stages. In the first stage, PointRCNN extracts the semantic features of each point, separates the front scenic spots, and extracts the preselection box. In the second stage, the candidate box is refined by integrating the semantic features and classification confidence of the first stage to obtain a more accurate bounding box. Finally, the experiments show that the average accuracy of the proposed method is the highest, with an average accuracy of 0.919 in the simple difficulty scenario, 0.897 in the medium difficulty scenario, and 0.839 in the difficult difficulty scenario, which are higher than those of the compared methods. Therefore, the proposed method can effectively improve the accuracy of power operation violation identification.

A Novel Intelligent Detection Algorithm of Aids to Navigation Based on Improved YOLOv4

Aiming at the problem of high-precision detection of AtoN (Aids to Navigation, AtoN) in the complex inland river environment, in the absence of sufficient AtoN image types to train classifiers, this paper proposes an automatic AtoN detection algorithm Aids-to-Navigation-YOLOv4 (AN-YOLOv4) based on improved YOLOv4 (You Only Look Once, Yolo). Firstly, aiming at the problem of an insufficient number of existing AtoN datasets, the Deep Convolutional Generative Adversarial Networks (DCGAN) is used to expand and enhance the AtoN image dataset. Then, aiming at the problem of small target recognition accuracy, the image pyramid is used to multi-scale zoom the dataset. Finally, the K-means clustering algorithm is used to correct the candidate box of AN-YOLOv4. The test on the test dataset shows that the improvement effect of AN-YOLOv4 is obvious. The accuracy rate of small targets is 92%, and the average accuracy (mAP) of eight different types of AtoN is 92%, which is 14% and 13% higher than the original YOLOv4, respectively. This research has important theoretical significance and reference value for the intelligent perception of the navigation environment under the intelligent shipping system.

Intelligent Tumor Cell Detection Method Based on Circulating Tumor Cell (CTC) Technology

&lt;p&gt;Cancer has become one of the greatest threats for human life. Doctors can get original images of the sick organs with the assistance of medical technology. However, manual interpretation of the original images is time consuming and labor consuming. Nowadays, the intelligent detection of tumor cell images is commonly adopted in cancer diagnosis. In this paper, we propose Imporved Selective Search (ISS) algorithm and CTCNet based on Circulating Tumor Cell (CTC) technology to improve the cancer images&amp;rsquo; detection efficiency. CTCs can be collected by a sampling needle with EpCAM antibody which can specifically bind to tumor cells. After fluorescent staining process, images obtained from sampling needle will be processed by the ISS algorithm for candidate region preselection. All of the eligible areas are evaluated with a self-designed neural network called CTCNet, resulting in efficient recognition of circulating tumor cells. During this process, all algorithms are accelerated through the GPU and NPU hardware platform, which further improves the detection speed of the system. In the experiment, we first verify the efficiency of proposed ISS algorithm by compared with Original Selective Search (OSS), we found that the number of candidate boxes reduced from 549 to 16 and the time consuming reduced by 0.3s after adopting the ISS algorithm. In order to evaluate the performance of the proposed 7-layer CTCNet, we compared CTCNet with SVM, BP neural network, AlexNet and VGGNet by using the dataset of 12312 samples from 30 patients, among them, there were 12 patients with early cancer and 18 patients with advanced cancer. And we got the highest recognition accuracy of 97.95% on CTCNet, which even beyond the VGGNet with deeper layers. In contrast to other combinations, we detect CTC in diverse clinical CTC images, the joint application of ISS algorithm and CTCNet achieves an outstanding system performance with accuracy up to 94.03% in the whole view images. Meanwhile, because of the application of a lightweight network in different hardware acceleration platform, the detection time of the CTC image in a single view can be less than 12s.&lt;/p&gt; &lt;p&gt;&amp;nbsp;&lt;/p&gt;

Detection of Missing Insulator Caps Based on Machine Learning and Morphological Detection.

Missing insulator caps are the key focus of transmission line inspection work. Insulators with a missing cap will experience decreased insulation and mechanical strength and cause transmission line safety accidents. As missing insulator caps often occur in glass and porcelain insulators, this paper proposes a detection method for missing insulator caps in these materials. First, according to the grayscale and color characteristics of these insulators, similar characteristic regions of the insulators are extracted from inspection images, and candidate boxes are generated based on these characteristic regions. Second, the images captured by these boxes are input into the classifier composed of SVM (Support Vector Machine) to identify and locate the insulators. The accuracy, recall and average accuracy of the classifier are all higher than 90%. Finally, this paper proposes a processing method based on the insulator morphology to determine whether an insulator cap is missing. The proposed method can also detect the number of remaining insulators, which can help power supply enterprises to evaluate the degree of insulator damage.

Collaborative Learning-Based Network for Weakly Supervised Remote Sensing Object Detection

Existing object detection algorithms rely excessively on instance-level labels, which are both time-consuming and expensive. In particular, for remote sensing images (RSI) with small and dense objects, the labeling cost is much higher than that of general images. Moreover, the propagation process of the labels over the noisy channel results in blurred and noisy information. To address the problem of obtaining RSI instance-level labels, we aim to propose a collaborative learning-based network for weakly supervised remote sensing object detection (CLN-RSOD). Compared with the state-of-the-art, the proposed model combines the advantages of two object detection sub-networks by jointly training. This improves the model's capability and thereby enhances the detection effect for multi-object in RSI. Moreover, we employ a mask-based proposal refinement algorithm for remote sensing images (MPR-RS) to optimize the candidate boxes. In addition, according to the data distribution characteristics of RSI, we introduce a new joint pooling module (JPM) in CLN-RSOD to enhance the backbone network's characterization of RSI. Finally, the experimental results on two public remote sensing datasets illustrate that the proposed weakly supervised learning method is superior to other weakly supervised methods and demonstrates the effectiveness of the proposed CLN-RSOD.