Detecting the Surface Defects of the Magnetic- Tile Based on Improved YOLACT ++

In the process of manufacture of magnetic tiles, especially in the production and transportation, the defects such as crack, stains, uneven and break will appear inevitably. To defect different defects of the magnetic tiles effectively, this paper proposes an improved YOLACT++ based on attention, and because of the YOLACT++ is a real-time instance segmentation model, which is constructed by deep convolution neural network. In our methods, we add attention mechanism SE-attention module and lightweight networks to the YOLACT++ structure, then use Soft-NMS (Non-Maximum Suppression) at the forecaster and add pre-trained weights. The experiments demonstrate that our method can achieve FPS gain of 3.9 points while mAP gain of 1.8 points over YOLACT++. The experimental results show that our method has fine robustness as well as stability for the detection of the magnetic tiles. The speed and accuracy of magnetic tile detection exceed the previous detection algorithm.