25 December 2023 in IEEE Access
Due to the large-scale changes of different forms of traffic signs and the rapid speed of vehicles, the detection accuracy and real-time performance of general object detectors are greatly challenged, especially the detection accuracy of small objects. In order to solve this problem, a multi-scale traffic sign detection model CR-YOLOv8 is proposed based on the latest YOLOv8. In the feature extraction stage, the attention module is introduced to enhance the channel and spatial features, so that the network can learn the key information of the small objects more easily. The RFB module is introduced in the feature fusion stage, which improves the feature diversity with less computational overhead and improves the network’s ability to detect multi-scale objects. By improving the loss function to enable the model to effectively balance multi-scale objectives during training, the model generalization ability is improved.The experimental results on TT100k dataset show that compared with the baseline network, the average detection accuracy of the improved method is increased by 2.3 %, and the detection accuracy of small objects is increased by 1.6 %, which effectively reduces the detection accuracy gap among different scales.
18 September 2023 in IEEE Access
The insulator defect targets in UAV (Unmanned Aerial Vehicle) images are often small and set against complex backgrounds. Consequently, traditional object detection algorithms commonly struggle to identify these minor defects. To enhance precision and recall in detecting insulator defects, a novel model, GC-YOLO (ghost convolution and centralized feature pyramid -You Only Look Once), based on YOLOv5s, has been introduced. GC-YOLO incorporates the Ghost convolution module in the backbone network, reducing feature redundancy and improving the inference speed of the feature extraction network. Moreover, an attention mechanism based on Coordinate Attention (CA) is integrated at the terminal of the backbone network, aimed at emphasizing the extraction of crucial information from target features. The Explicit Visual Center Block (EVCBlock) module from Centralized Feature Pyramid Network (CFPNet) is introduced in the neck layer to effectively fuse multi-scale features and enhance the feature map’s characterization capability. Furthermore, in order to enhance the precision in detecting small-sized defects, a small object detection head is also added to the detection layer based on CFPNet. Experimental results demonstrate that GC-YOLO achieves a recall of 89.7% and mAP@0.5 of 94.2%, surpassing YOLOv5s by 7% and 6.5%, respectively. The proposed algorithm exhibits superior detection precision in complex scenes, providing a theoretical basis for intelligent and mechanized railway monitoring systems.