A sea cucumber recognition network based on improved YOLOv5

A lightweight network based on YOLOv5 is proposed in this paper to improve the real-time detection ability of underwater targets for fishing gear and to solve the difficulty of deploying model algorithms on embedded devices. First, the Shuffle_Block module replaces the leading feature extraction network in YOLOv5, reducing parameters and improving the algorithm's inference speed. Second, this module is combined with depthwise separable convolution to construct the feature fusion Shuffle-PANet, significantly reducing network parameters and improving detection speed while ensuring accuracy. The proposed method in this paper has been verified to reduce the parameter count by 89% compared to the YOLOv5 source code while doubling the detection speed of the source code. Additionally, the weight file size is reduced by 83%. The mAP50 reaches 96.3%, which is only a 2% decrease compared to YOLOv5. The lightweight network proposed in this paper can recognize sea cucumbers well and has fast recognition speed and lightweight design characteristics. Shuffle-YOLOv5 has significant advantages compared to the original model and can complete real-time target detection on low-power embedded devices.