An Underwater Dense Small Object Detection Model Based on YOLOv5-CFDSDSE

Abstract

Underwater target detection is a key technology in the process of exploring and developing the ocean. Because underwater targets are often very dense, mutually occluded, and affected by light, the detection objects are often unclear, and so, underwater target detection technology faces unique challenges. In order to improve the performance of underwater target detection, this paper proposed a new target detection model YOLOv5-FCDSDSE based on YOLOv5s. In this model, the CFnet (efficient fusion of C3 and FasterNet structure) structure was used to optimize the network structure of the YOLOv5, which improved the model’s accuracy while reducing the number of parameters. Then, Dyhead technology was adopted to achieve better scale perception, space perception, and task perception. In addition, the small object detection (SD) layer was added to combine feature information from different scales effectively, retain more detailed information, and improve the detection ability of small objects. Finally, the attention mechanism squeeze and excitation (SE) was introduced to enhance the feature extraction ability of the model. This paper used the self-made underwater small object dataset URPC_UODD for comparison and ablation experiments. The experimental results showed that the accuracy of the model proposed in this paper was better than the original YOLOv5s and other baseline models in the underwater dense small object detection task, and the number of parameters was also reduced compared to YOLOv5s. Therefore, YOLOv5-FCDSDSE was an innovative solution for underwater target detection tasks.

BPN-YOLO: A Novel Method for Wood Defect Detection Based on YOLOv7

“…Therefore, to reduce the computational effect as well as to improve the performance, we introduce a convolution operation called partial convolution [32]. Pconv further optimizes the size of the computation and reduces the number of FLOPs by exploiting the redundancy of the feature map [33]. In this way, only some of the input channels use regular convolution to extract features without affecting the other channels, which not only reduces the computational redundancy and memory access but also improves inference speed and detection accuracy.…”