Retail Commodity Image Recognition based on WS-DAN

Abstract

Computer vision technology has landed very widely in daily life, business, security, and other fields, especially the application represented by image recognition technology has penetrated into all aspects of life. In the business scenario of brand owners, the efficiency and convenience brought by commodity recognition technology is also particularly prominent, which can bring cost reduction and efficiency increase for enterprises. Especially in the retail goods industry, the use of image recognition technology can greatly improve productivity levels. However, the number of retail goods categories is huge. The differences in the packaging of similar goods may be very subtle, such as the difference in logo position, the difference in the label text, or even the difference in the color of a small area. These differences are difficult to be accurately identified in general image recognition. Ordinary image recognition technology cannot meet the demand, so it is necessary to apply fine-grained image recognition technology. In this paper, an mechanism of attention to guide data augmentation fine-grained neural networks image recognition, i.e. WS-DAN (Weakly Supervised Data Augmentation Network) is used for fine-grained recognition of retail goods. The model was trained and validated using the dataset provided by Baidu Al (Artificial Intelligence) studio, and experiments were carried out to investigate the impact of various factors on image recognition accuracy and the difficulty in recognizing products by category.