Lightweight Sea Cucumber Recognition Network Using Improved YOLOv5

Aiming at the problems of the poor real-time detection ability of intelligent devices for underwater targets and the difficulty of deploying complex model algorithms on embedded devices with limited resources, this paper proposed a lightweight feature extraction module GGS, which combines traditional algorithm downsampling with deep separable convolution downsampling and introduces a parameter-free attention mechanism to extract features from input data at multiple scales and focus on target information. The paper constructs the GGS-PF-YOLOv5 network by replacing the backbone extraction network in YOLOv5 with the GGS module and the GGS-PANET by combining the GGS module with depth separable convolution for feature fusion, and the PfAAMLayer non-parameter attention mechanism is introduced to enhance the feature extraction capabilities of the model by focusing on the feature information of the two dimensions of channel and space, improving the identification accuracy of sea cucumbers while keeping the number of network parameters low. Experimental results show that the proposed network, with remarkably few network parameters, outperforms the original YOLOv5 model regarding detection speed while achieving comparable accuracy. The GGS-PF-YOLOv5 model reduces the parameter volume by 94% compared to YOLOv5s source code while doubling detection speed, with a weight file size of only 1.08M, 92% smaller than the source code. The (Map50) only decreases by 1.5% compared to YOLOv5s, which indicates that this paper proposed model can achieve real-time target detection on low-power embedded devices while maintaining high levels of accuracy.