Multi-scale Defect Detection Network for Tire Visual Inspection

Automatic defect detection in industrial inspection remains a challenging task due to the diversity of products and the complexity of textures. In this work, we focus on the detection task in tire industry and develop a multi-scale defect detection network (MDDN), which captures both semantic and texture features using two parallel deep convolutional network. High-level semantic features containing defect shapes and locations are captured via a semantic-aware network, simplified by an off-the-shelf fully convolutional network. A simple yet effective texture-aware network is simultaneously developed to complement the details filtered out by the sub-sampling. Pixel-wise detection results are then obtained by integrating features with semantic and texture information. Moreover, we carefully designed a multi-scale preprocessing strategy to make the model describe defects more accurately with the help of the texture similarity in the tire image. Extensive experiments demonstrate that MDDN can achieve significant performance improvement in detecting defects with different sizes over some existing methods.