A lightweight network based on SCMYOLO for accurate and efficient underwater fish detection

Abstract

The accurate and rapid identification of fish in natural underwater settings is crucial for the intelligent monitoring of fisheries. However, the computational requirements of intricate models necessitate advanced hardware capabilities, limiting the deployment of fish identification models within underwater monitoring systems. To address this, we introduce a lightweight fish object detection model named SCMYOLO. This model employs several innovations to reduce complexity and enhance performance. Firstly, we replace the yolov5 backbone network with MobileNetV3, significantly reducing network parameters and the computational load as measured by floating-point operations. Secondly, we integrate the high-level screening-feature fusion pyramid structure with the Neck layer of yolov5 and combine the C3 module with the context guided network (CGNetwork) module. This approach improves the extraction of low-level features and addresses challenges associated with multi-scale object detection, thereby enabling lightweight multi-scale feature extraction and fusion. Lastly, we incorporate the SIoU loss function to refine the accuracy of bounding box predictions. Our experimental findings demonstrate that the SCMYOLO model achieves an 88.4% reduction in model parameters and an 86.2% decrease in floating-point operations, with the size of the model file constituting just 14.6% of the original. Furthermore, the model sustains high-performance metrics with precision, recall, and mAP@0.5 all exceeding 90%, indicating robust multi-scale detection capabilities. Given these attributes, SCMYOLO is particularly suited for deployment on various edge computing devices, offering an efficient solution for fish detection and identification in complex underwater environments. Source code is available at https://github.com/systemavailable/SCMYOLO.