The work aims to fast count the product information on shelves, and propose an automatic recognition method of shelf products based on deep neural network. The image of the shelf products collected by the camera was processed by the deep neural network algorithm to obtain the SKU and position of the products in the image. Aiming at the dense detection scenario of shelf product recognition, this method improved the general deep neural network object detection algorithm: the algorithm was divided into two stages of detection and classification and a part of the network structure was redesigned. At the same time, this method was compared with the traditional shelf product recognition methods and the general deep neural network object detection methods. From the experiment results, the average precision of the model reached 96.5% in the detection stage and 99% in the classification stage. In whole image test, the precision was 98.17% and the recall was 97.05%. Compared with prior works by traditional object detection methods for product recognition on shelves or SIFT artificial operators to extract features and classify product SKU, this method greatly improves the detection rate and classification precision rate, which has good application potential.