| Summary: | The lateral line scales of fish are an important phenotype of fish species. As an important countable feature, the accurate and effective counting of lateral line scales is an important reference standard for breeding, determining the growth status of fish, and identifying fish species. At present, the statistical work of fish lateral line scales mainly depends on manual statistics and semi-automatic methods, which cannot meet the current needs of sustainable development and precision digital fisheries. The method based on computer vision and deep learning can provide an a real-time, efficient and non-contact method for identifying and counting fish lateral line scales. However, it is still a challenge due to the high similarity between fish scales and the variable size issues caused by the free movement of the fish. Hence, we proposed a transformer module improved YOLOv5 model (TRH-YOLOv5) for fish lateral line scale detection and counting, which focus on the high similarity of fish scales. In addition, we design a small target detection module in the head layer to address the challenge of multi-scale fish. To evaluate the effective of our method, performance of proposed model is analyzed on different type fish dataset and it is also compared with classical method including SSD, YOLOv4 and YOLOv5. Comprehensive experimental results show that the proposed model achieves fine results (e.g., 98.8% precision, 96.7% recall and 99.0% mean average precision) with relatively lower computational coat (e.g., 16.1M model size) and fast detection speed (e.g., 37 FPS) compared with the benchmark algorithm. The TRH-YOLOv5 model is also used for swimming fish video to detect fish lateral line in real-time and can be integrated into aquaculture vision system for aquaculture precision and sustainable management.
|
|---|