Apple rapid recognition and processing method based on an improved version of YOLOv5

To improve the processing efficiency of apple picking robots, we have developed a rapid apple recognition and tracking method based on an improved version of the YOLOv5 algorithm. In the algorithm, the backbone standard convolution module in YOLOv5 is replaced by the inverted residual convolution module in the light-weight MobileNetv2 network. Additionally, the least-squares method is introduced to correct the misjudged data output results in the model, which is more suitable for identifying various apple shapes. Furthermore, by introducing the method of target association recognition, the multitarget picking path is designed according to the correlation between the confidence levels of the identified targets. The combination of the above method in improving YOLOv5s can reduce the size of the model and improve the detection speed, which is convenient for the migration and application of the model in hardware devices. The volume of the improved model is 6.01 Mb, which is 57% smaller than that of the original model, and the speed of data processing is increased by 27.6%, which reaches 90 frames per second. Our results show that the method of target association recognition reduces the processing time of the model selection process and improves the efficiency by 89%. Compared with seven other deep networks, such as YOLOv8s, the improved YOLOv5 model has the fastest speed when detecting apples while maintaining better detection accuracy. The improved YOLOv5 model shows a good detection effect, which provides a technical basis for the detection of fruits picked by apple picking robots