Global Context Network for Steel Surface Defect Detection

Surface defect detection has been spotlighted in the product quality control. There are lots of methods focused on the handcrafted optical features and have worked well under specified conditions. However, effectively detecting defects in products is nontrivial. Among the challenge is the complexity of surface defect, such as micro defect with noise, at vastly different scales. In order tackle these problems, we propose a feature fusion network using global context block for surface defect detection. A pipeline is presented that evaluates defect images with 300×300 resolution. In the framework, the global context block is relined, which fuses information effectively between different feature maps. Experimental results on steel defect datasets prove that our approach yields scores of map > 0.6 for all surface defects and provides a remarkably fast test speed, at 20 frames per second.