Visual Recognition of Traffic Signs in Natural Scenes Based on Improved RetinaNet

Aiming at recognizing small proportion, blurred and complex traffic sign in natural scenes, a traffic sign detection method based on RetinaNet-NeXt is proposed. First, to ensure the quality of dataset, the data were cleaned and enhanced to denoise. Secondly, a novel backbone network ResNeXt was employed to improve the detection accuracy and effection of RetinaNet. Finally, transfer learning and group normalization were adopted to accelerate our network training. Experimental results show that the precision, recall and mAP of our method, compared with the original RetinaNet, are improved by 9.08%, 9.09% and 7.32%, respectively. Our method can be effectively applied to traffic sign detection.

A Decision Support System for Power Components Based on Improved YOLOv4-Tiny

The traditional image object detection algorithm applied in power inspection cannot effectively position power components, and the accuracy of recognition is low in scenes with some interference. In this research, we proposed a data-driven power detection method based on the improved YOLOv4-tiny model, which combined the ResNet-D module and the adjusted Res-CBAM to the backbone network of the existing YOLOv4-tiny module. We replaced the CSPOSANet module in the YOLOv4-tiny backbone network with the ResNet-D module to reduce the FLOPS required by the model. At the same time, the adjusted Res-CBAM whose feature fusion ways were replaced with stacking in the channels was combined as an auxiliary classifier. Finally, the features of five different receptive scales were used for prediction, and the display of the results was optimized by merging the prediction boxes. In the experiment, 57134 images collected on the power inspection line were processed and labeled, and the default anchor boxes were re-clustered, and the speed and accuracy of the model were evaluated by video and validation set of 3459 images. Processing multiple pictures and videos collected from the power inspection projects, we re-clustered the default anchor box and tested the speed and accuracy of the model. The results show that compared with the original YOLOv4-tiny model, the accuracy of our method that can position objects under occlusion and complex lighting conditions is guaranteed while the detection speed is about 13% faster.

Multiple Instance Learning Convolutional Neural Networks for Fine-Grained Aircraft Recognition

The key to fine-grained aircraft recognition is discovering the subtle traits that can distinguish different subcategories. Early approaches leverage part annotations of fine-grained objects to derive rich representations. However, manual labeling part information is cumbersome. In response to this issue, previous CNN-based methods reuse the backbone network to extract part-discrimination features, the inference process of which consumes much time. Therefore, we introduce generalized multiple instance learning (MIL) into fine-grained recognition. In generalized MIL, an aircraft is assumed to consist of multiple instances (such as head, tail, and body). Firstly, instance-level representations are obtained by the feature extractor and instance conversion component. Secondly, the obtained instance features are scored by an MIL classifier, which can yield high-level part semantics. Finally, a fine-grained object label is inferred by a MIL pooling function that aggregates multiple instance scores. The proposed approach is trained end-to-end without part annotations and complex location networks. Experimental evidence is conducted to prove the feasibility and effectiveness of our approach on combined aircraft images (CAIs).

Deep Learning for Automatic Image Segmentation in Stomatology and Its Clinical Application

Deep learning has become an active research topic in the field of medical image analysis. In particular, for the automatic segmentation of stomatological images, great advances have been made in segmentation performance. In this paper, we systematically reviewed the recent literature on segmentation methods for stomatological images based on deep learning, and their clinical applications. We categorized them into different tasks and analyze their advantages and disadvantages. The main categories that we explored were the data sources, backbone network, and task formulation. We categorized data sources into panoramic radiography, dental X-rays, cone-beam computed tomography, multi-slice spiral computed tomography, and methods based on intraoral scan images. For the backbone network, we distinguished methods based on convolutional neural networks from those based on transformers. We divided task formulations into semantic segmentation tasks and instance segmentation tasks. Toward the end of the paper, we discussed the challenges and provide several directions for further research on the automatic segmentation of stomatological images.

Wheel fault diagnosis model based on multichannel attention and supervised contrastive learning

As wheels are important components of train operation, diagnosing and predicting wheel faults are essential to ensure the reliability of rail transit. Currently, the existing studies always separately deal with two main types of wheel faults, namely wheel radius difference and wheel flat, even though they are both reflected by wheel radius changes. Moreover, traditional diagnostic methods, such as mechanical methods or a combination of data analysis methods, have limited abilities to efficiently extract data features. Deep learning models have become useful tools to automatically learn features from raw vibration signals. However, research on improving the feature-learning capabilities of models under noise interference to yield higher wheel diagnostic accuracies has not yet been conducted. In this paper, a unified training framework with the same model architecture and loss function is established for two homologous wheel faults. After selecting deep residual networks (ResNets) as the backbone network to build the model, we add the squeeze and excitation (SE) module based on a multichannel attention mechanism to the backbone network to learn the global relationships among feature channels. Then the influence of noise interference features is reduced while the extraction of useful information features is enhanced, leading to the improved feature-learning ability of ResNet. To further obtain effective feature representation using the model, we introduce supervised contrastive loss (SCL) on the basis of ResNet + SE to enlarge the feature distances of different fault classes through a comparison between positive and negative examples under label supervision to obtain a better class differentiation and higher diagnostic accuracy. We also complete a regression task to predict the fault degrees of wheel radius difference and wheel flat without changing the network architecture. The extensive experimental results show that the proposed model has a high accuracy in diagnosing and predicting two types of wheel faults.

Vehicle and pedestrian target detection in auto driving scene

To improve the target detection accuracy and speed of autonomous driving in various weather environments and small target traffic senarios,an improved YOLOV4 target detection model based on CSPDarknet45_G backbone network is proposed in this paper.By adding a new DBG module which consists of DArknetConv2D + BN + GELU activation function,this model is enhanced in generalization ability and accuracy. We also improved Res unit residual module to enhance shallow features fusing with deep feathers and reduced the number of neurons in the CSP module to simplify the module structure.The K-Means++ clustering algorithm is introduced to obtain the size of the prior box used for target detection to satisfy the data set in this paper. In the captured target vehicle image data set, the model detection result shows that the improved YOLOV4 model achieve an average detection accuracy of 90.45%, a recall of 94.37%, and an FPS of 50 frames per second when the IOU is taken as 0.5, which meet the real-time and accuracy of the detection task in this paper.

Understanding Traffic Patterns of Covid-19 IoC in Huge Academic Backbone Network SINET

Recently, APT (Advanced Persistent Threats) groups are using the COVID-19 pandemic as part of their cyber operations. In response to cyber threat actors, IoCs (Indicators of Compromise) are being provided to help us take some countermeasures. In this paper, we analyse how the coronavirus-based cyber attack unfolded on the academic infrastructure network SINET (The Science Information Network) based on the passive measurement with IoC. SINET is Japan's academic information infrastructure network. To extract and analyze the traffic patterns of the COVID-19 attacker group, we implemented a data flow pipeline for handling huge session traffic data observed on SINET. The data flow pipeline provides three functions: (1) identification the direction of the traffic, (2) filtering the port numbers, and (3) generation of the time series data. From the output of our pipeline, it is clear that the attacker's traffic can be broken down into several patterns. To name a few, we have witnessed (1) huge burstiness (port 25: FTP and high port applications), (3) diurnal patterns (port 443: SSL), and (3) periodic patterns with low amplitude (port 25: SMTP) We can conclude that some unveiled patterns by our pipeline are informative to handling security operations of the academic backbone network. Particularly, we have found burstiness of high port and unknown applications with the number of session data ranging from 10,000 to 35,000. For understanding the traffic patterns on SINET, our data flow pipeline can utilize any IoC based on the list of IP address for traffic ingress/egress identification and port filtering.

Research on Lightweight Citrus Flowering Rate Statistical Model Combined with Anchor Frame Clustering Optimization

At present, learning-based citrus blossom recognition models based on deep learning are highly complicated and have a large number of parameters. In order to estimate citrus flower quantities in natural orchards, this study proposes a lightweight citrus flower recognition model based on improved YOLOv4. In order to compress the backbone network, we utilize MobileNetv3 as a feature extractor, combined with deep separable convolution for further acceleration. The Cutout data enhancement method is also introduced to simulate citrus in nature for data enhancement. The test results show that the improved model has an mAP of 84.84%, 22% smaller than that of YOLOv4, and approximately two times faster. Compared with the Faster R-CNN, the improved citrus flower rate statistical model proposed in this study has the advantages of less memory usage and fast detection speed under the premise of ensuring a certain accuracy. Therefore, our solution can be used as a reference for the edge detection of citrus flowering.