Research on commodity recognition method based on YOLOv8_SMB

In order to solve the problems of low accuracy, missed detection and false detection caused by the diverse and large distribution differences of retail commodity image data in different environments, a YOLOv8_SMB-based commodity recognition and detection method was proposed. Firstly, the dataset was enriched by data augmentation of a single commodity image, and the diversity of the data was enhanced by the method of mixup interpolation to solve the problem of small samples. Secondly, the DIOU loss function is used to minimize the distance between the two target frames, so that the convergence speed and regression speed are improved. Then, the feature fusion was carried out by introducing three modules combining SDI, MDFM and BiFPN to effectively improve the feature representation ability. Then, MLKA was used to enhance the quality of the extracted features by fusing large kernel decomposition and multi-scale learning strategies, which significantly improved the accuracy of the model. Finally, DCNv3 and Faster modules are introduced into C2f to improve the computing speed and accuracy without increasing the amount of computation. The results show that the accuracy of the YOLOv8_SMB in the commodity recognition model proposed in this paper can reach 93.9% and 91.1% mAP@50, which are 2.6% and 20% higher than the original model, respectively. The results show that the accuracy of the YOLOv8_SMB in the commodity recognition model proposed in this paper can reach 93.9% and 91.1% mAP@50, which are 2.6% and 20% higher than the original model, respectively.