Abstract: Aimed at the problems of low real-time commodity detection, poor detection of stacked commodities, and difficulties in classifying similar commodities in the intelligent settlement task of unmanned supermarkets, a commodity recognition algorithm based on improved YOLOv4 is proposed. The algorithm used the lightweight network MobileNetv2 for feature extraction to speed up the detection speed. The channel attention and spatial attention were introduced into the inverted residual structure of MobileNetv2 to amplify the local feature weight, and thus enhancing the detection ability of stacked commodities. Focal loss was used in the loss function to solve the difficult classification problem with small inter-class difference. The experimental results show that this method achieves 80.3% accuracy and a detection speed of 73 FPS on self-built product data sets, which is better than the YOLOv4 algorithm.