Detection of fresh tidiness in supermarket: a deep learning based approach

Ying Zang, Chenglong Fu, Qingshan Liu, Shuguang Zhao, Chaotao Ding

Abstract: Psychological research shows that the tidiness of fresh display in supermarkets will affect consumers’desire to buy to a great extent. Therefore, whether the supermarket administrator can organize the shelves in time and keep the fresh on the shelves neatly placed has a great impact on their sales performance. In order to solve this problem, we use the method of calculating the distance between the neatly placed pictures and the texture feature vector of the collected test images to measure the similarity of the two pictures, so as to determine whether the goods are neatly placed. HistNet is a texture classification network which takes ResNet as the baseline network and introduces histogram layer, which can classify images with texture features very well. Based on HistNet, we use large margin cosine loss(LMC loss) and CrossEntropy loss for joint training to improve the discrimination ability of the network by increasing inter-class differences and reducing intra-class differences. Then change the convolution method of learning bin center in the histogram, make the texture features learned by the histogram more abundant, and add convolutional block attention module(CBAM) attention before the first convolution in the histogram layer to further improve the network performance. On the basis of the optimized network, we remove the last layer and measure the extracted texture features directly. We do a large number of experiments on a variety of distance measurement methods and choose cosine distance to judge the tidiness of fresh display in the supermarket. We tested the improved algorithm on DTD, MINC-2500 and GTOS-mobile datasets, and the accuracy reached 71.79%, 83.21% and 81.24%, respectively. The optimized network improves the accuracy of texture classification. In the application of supermarket fresh tidiness detection, it can well distinguish between tidiness and confusion, and has a broad application prospect.