Aiming at the target detection algorithm for commodity detection accuracy is not high, stacked commodity detection ability is low, commodity classification ability is poor and other problems. Construct an improved algorithm model OCR-YOLOv8 based on YOLOV8n for commodity recognition algorithm. First, the use of full-dimensional dynamic convolution for the picture for feature extraction, In contrast to convolutional neural networks, this method can utilise four spaces of attentional learning, and have a stronger understanding of the picture content. Second, the C2fSTR attention module is added to the backbone network in order to capture rich image information at different scales and realize the exchange of information between different pictures, which in turn enables the backbone network to improve its classification ability for complex features. Finally, the RFB receptive field network is added in front of the three prediction heads of YOLOv8n to improve the accuracy of predicting items by simulating human vision.