Marine zoobenthos recognition algorithm based on improved lightweight YOLOv5

Abstract

Detecting the distribution and density of marine zoobenthos is crucial for monitoring healthy coastal ecosystems and for growth reference tracking in precision aquaculture. However, current detection algorithms for marine zoobenthos have high computational complexity and cannot guarantee a balance between accuracy and speed, limiting their deployment in fishery equipment. This study used a portion of the Augmented Underwater Detection Dataset, a large underwater biological dataset containing marine zoobenthos data. A marine zoobenthos recognition algorithm was proposed for sea cucumbers, sea urchins, and scallops based on an improved lightweight YOLOv5, which can recognize the three types of marine zoobenthos. In the image enhancement module, an underwater image enhancement algorithm based on color balance and multi-input fusion is used, which turns the blurred image into a natural appearance of the seabed image. The lightweight backbone network EfficientnetV2-S was chosen to replace the original YOLOv5 backbone network, reducing network parameter calculations and improving recognition speed. A Bottleneck Transformer was introduced into the backbone network, and an attention mechanism based on the convolution module was introduced to construct the embedded Convolutional Block Attention Module in the Neck structure of YOLOv5, thereby improving the recognition accuracy of the lightweight YOLOv5 model. The experimental results showed that the mAP of the proposed algorithm reached 0.941, which is an improvement of 0.002 compared with the original YOLOv5l algorithm. The computation of this algorithm is 37.0 FLOPs (G), the model size is 54 MB, and the inference time is 5.9 ms. Compared to the original YOLOv5l algorithm, the reductions are 66.1%, 40.5%, and 39.2%. The proposed algorithm efficiently identified and classified marine zoobenthos.