Optimization of Tiny YOLOv7 Lightweight Model Through Network Structure Re-parameterization for Coral Reef Fish Identification

Coral reefs face significant threats, leading to rapid declines in reef fish populations due to destructive fishing activities and climate change. Conventional methods used to estimate coral reef fish biomass abundance are underwater visual census (UVC) and diver-operated video (DOV). Limitations of underwater visual census (UVS) and diver-operated video (DOV) methods include potential bias from the presence of the diver, reliance on diver skill and video quality, time-consuming analysis, and high-quality equipment, all of which can lead to inaccuracies in identification. This study addresses these challenges by developing a cost-effective reef fish monitoring system using an IoT platform equipped with stereo cameras. This system is designed as edge computing, object detection is carried out on the IoT platform to overcome time-consuming analysis. The system detects fish species and measures the length of the fish using depth information from the stereo camera. To improve detection accuracy and speed, we use a reparametrized YOLO Tiny single-stage CNN architecture. By optimizing the model with CBL and CBM blocks, the Tiny YOLOv7 with adjusting the number of CNN layer achieved an F1 score of 86.64 and a mean average precision (mAP) of 90.50%. The scaled-down model efficiently uses computing resources, operating at only 3 % of the embedded system capacity, equivalent to 472 GFLOPs. The results of the embedded transfer learning model obtained a detection speed performance of 11 FPS from a 24 FPS camera stream.