Machine Vision-Based Recognition for fruit cracking in cherry

In order to realize the intelligent real-time and accurate identification of cherry fruit cracking, a set of end-to-end data remote acquisition device was designed with normal and cracked cherry fruit as the research object. Aiming at the problem of unbalanced number of cherry cracked fruit and normal fruit samples, a sample intelligent transplantation data enhancement algorithm is proposed. At the same time, the YOLOX-S model is improved and designed. CBAM-MASFF-YOLOX increases the output of the feature map in the backbone feature extraction network, and introduces the CBAM attention mechanism to ensure that the high-level semantic information is not lost and the feature extraction is strengthened. In the feature pyramid module, after feature extraction of the newly transmitted feature map, it is introduced into the MASFF multilevel adaptive feature fusion module together with the original three feature maps. Finally, the binary cross entropy in the loss function is replaced by VariFocalLoss, and IoU Loss is replaced by α-IoU Loss to improve the training weight and robustness of the model for cherry samples. The results of the model test show that the average accuracy of the whole category of cherry recognition reaches 94.18 %, and the detection time is 34.29 ms / sheet. Compared with the original YOLOX-S, the average accuracy of normal and cracked cherry fruit was increased by 4.07 % and 7.47 %, respectively, and the mAP was increased by 5.77 %.The method proposed in this study has higher accuracy while satisfying the application of detection time.