Research on Fresh Squid Body Recognition Based on Improved YOLO v5 Model

In order to improve the intelligence level and efficiency of squid fishing, a new target detection and recognition model, YOLO v5-SQ based on improved YOLO v5, is proposed based on the operation process of squid fishing industry and the sorting characteristics of the squid fishing industry. Based on the stiffer cartilage of squid, which makes its body relatively straight, the length of the body is innovatively measured and graded to synchronize the measurement and grading of fresh squid during the sorting process, so as to improve work efficiency. The improved model YOLO v5-SQ uses image processing to increase sample size and improve learning effect. To better capture the details of squid neck, ConvNeXt is used to preserve the features of small details. To solve the problem of squid posture diversity and large changes in neck characteristics, BiFPN(Bi-directional Feature Pyramid Network) structure is used in the part of network feature fusion, which can enhance the network’s ability to fuse multi-scale target features. During squid transport, if no posture correction is made, the body may not be parallel to the final label frame, resulting in too large difference between the converted size and the actual size. Therefore, a mechanical structure is used to modify the squid posture at the front of recognition, so that the middle line of the body is as parallel as possible to the horizontal line of the shooting interface and to reduce errors. The results show that the accuracy of YOLO v5-SQ detection can reach 98.7%, the average error of body length is 1.57%, and the speed of time detection when calling the camera is 47.83 f/s. This method can achieve the expected effect and has high practical value. It also provides some improvement ideas for the intelligent research of other seafood operations.