Surface Defect Detection of Frozen Dumplings Based on Improved U-Net Network

In order to realize automatic detection of surface defects of frozen dumplings, a U-shaped semantic segmentation ResNet_Unet based on U-Net is proposed in this paper. ResNet50 is used as the encoder of the semantic segmentation network to enhance the feature extraction capability. In order to enhance the recognition accuracy of small targets, expanding the receptive field of the multi-scale output feature map of the encoder is proposed. And embedding BN normalizes the data before decoding module convolution, which can speed up training and improve generalization of the model. In addition, the label is smoothed when the loss is calculated to prevent the model from placing too much faith in the predictions. The experimental result shows that the mPA and mIoU of the ResNet_Unet network are 83.20% and 80.44%, respectively. This study provides a reference for intelligent segmentation of surface defects in frozen dumplings.