BiDNet: A Real-Time Semantic Segmentation Network With Antifeature Interference and Detail Recovery for Industrial Defects

In recent years, there has been an increasing demand for surface defect detection with the development of intelligent manufacturing. The semantic segmentation is suitable for achieving precise and intelligent surface defect detection. However, three issues prevail in current surface defect segmentation methods: feature interference, detail missing, and computationally expensive. To address these limitations, we propose the bilateral decoder network (BiDNet), a novel real-time semantic segmentation framework with shallow and deep branches. Computationally expensive is due to the high resolution of the shallow feature maps. To solve this problem, BiDNet uses shallow branches for shallow feature maps with a high resolution to ensure speed and deep branches for deep feature maps with a low resolution to guarantee accuracy. Feature interference is caused by the direct fusion of deep feature maps of different sizes. To solve this problem, we propose a multiscale feature channel attention (MFCA) mechanism to compute the contribution of feature maps from different layers and accordingly fuse them better. The detail missing is due to the gradual downsampling in the encoder stage. To solve this problem, we propose a multiscale feature spatial attention (MFSA) mechanism to compute the importance of each position of the feature map for different branches to recover the details better. Extensive experiments on mobile phone screen surface defect (MSD), magnetic tile defect (MTD), and our glass surface defect (GSD) dataset show that our performance consistently outperforms the state of the art. The code is available at: https://github.com/jiaweipan997/BiDNet.