Traditional fish body directional conveyance suffers from high labor intensity and low efficiency in the fish processing process. This study employed machine vision technology to achieve directional identification of fish bodies for a special self-developed device. 2000 images of fish in four directions are selected to establish a database. The data set is finally expanded to 3000 images through image processing. A YOLOv5s object detection model was used to determine the head-tail and dorsal-ventral directions of silver carp by analyzing their morphological features and physical properties to achieve directional arrangement of the fish bodies for the device. The identification results were utilized to manage the fish body direction and successfully accomplished head-tail and dorsal-ventral directional arrangement of the fish. The detection model demonstrated an accuracy of 99.76%, a recall rate of 99.59%, an average precision of 99.5%, and a F1-score of 99.66% on the test dataset. The model is compared with other models in the YOLOv5 series as well as the YOLOv8 model. It is found that the accuracy and recall rates are not much different, but the YOLOv5s model used in this paper is only 14.1MB in size, and the average detection speed of the model is as high as 0.029s. At the same time, the performance test of the device was carried out. The results indicated at the conveying speeds of 0.05, 0.45, and 0.6 m/s respectively for the lifting system, separation system, and direction identification and assignment system, the average success rate could reach 97.2% for head-tail directional arrangement and 95.6% for dorsal-ventral directional arrangement. Additionally, the device achieved a directional conveying throughput of up to 15 fish per minute. These findings may facilitate the design of conveyor devices for the arrangement of fish bodies in freshwater fish processing industry.

Silver Carp, Head-Tail Orientation, Dorsal-Ventral Orientation, Object Detection, YOLOv5s