Robust detection of farmed fish by fusing YOLOv5 with DCM and ATM

Abstract

Accurate object detection of farmed fish is vital for fish counting and measuring. However, the blurring, occlusion, and varying density of underwater fish images from aquaculture ponds result in low precision and poor robustness. This paper proposes a robust detection model called DCM-ATM-YOLOv5, in which the deformable convolution module (DCM) and adaptive threshold module (ATM) are integrated into YOLOv5. The farmed fish detection model is based on YOLOv5, considering its high precision and real-time capabilities. To achieve higher precision on blurred targets, the DCM is applied to offset the sampling locations and to focus more on fish so that the features of blurry fish are enhanced and the background is suppressed. The ATM is introduced to alleviate missed detections resulting from fixed thresholds by generating appropriate thresholds for different densities and predicting the thresholds dynamically, improving the robustness of the proposed model. Two experiments based on data from aquaculture ponds were designed to test the performance of the proposed model. The average precision and recall of DCM-ATM-YOLOv5 were 97.53% and 98.09%, respectively. Compared with the average precision and recall of YOLOv5, those of DCM-ATM-YOLOv5 were improved by 1.92% and 1.97%, respectively. The robustness of the proposed model was evaluated based on data with different densities and settings. The detection results obtained from datasets of varying densities and settings demonstrated that the detection performance of DCM-ATM-YOLOv5 was still superior to that of YOLOv5, indicating that the proposed model has significant stability and robustness. The experimental results showed that DCM-ATM-YOLOv5 has high precision and strong robustness in genuine aquaculture settings and can provide a feasible solution for the detection of farmed fish.