MR-UNet Commodity Semantic Segmentation Based on Transfer Learning

Abstract

The display of commodity in the e-commerce field requires a large amount of labor to carry out image cutting. How to pick out the interested objects quickly and accurately from the images of commodity is an arduous challenge for commodity segmentation. Although the development of neural network has improved the accuracy of commodity segmentation, false negative results are still inevitable. In order to promote the performance of commodity segmentation, an UNet-based neural network, MR-UNet (more-residual UNet), for commodity segmentation is proposed. By improving the encoder of UNet, deepening the network layers of the encoder and adding residual blocks, the MR-Module is constructed to enhance the ability of the model to extract details of goods. In order to reduce the checkerboard effect, bilinear interpolation is utilized instead of deconvolution in UNet model. By migrating the weight of ResNet and generalizing the model on the enhanced data set, MR-UNet shows better performance in commodity semantic segmentation. The experiments were carried out on the labeled Fashion Product Images Dataset, and the results show that MR-UNet model acquired outstanding results in segmenting hollow commodities. Compared with several popular deep learning models UNet, PSPNet, FCN, DeepResUNet, UNet3+ in terms of commodity semantic segmentation, MR-UNet performs better in all indicators such as Dice Coefficient, MIoU, Matthews Coefficient, and Pixel Accuracy.