Fully Convolutional Networks For Surface Defect Inspection

Surface defect inspection,which assesses the quality of product,acts as an important role in industrial environment.With the development of computer vision,automated computer visual inspection is currently the main form that improves the industrial automation level significantly.However most of the state-of-the-art computer vision based defect inspection algorithms are pattern-based approaches.Feature extractor designing requires designers to have rich prior knowledge,and the challenge is that such methods can hardly be generalized or reused and may be inapplicable to the image conditions found in real runnel images.The special pre-and post-processing and the caseby-case feature extractor-designing make the development cycle relative complex and time-consuming.Therefore,the goal of this paper is to design a general and pixel-wise algorithm framework for surface defect inspection in industrial environment.Aiming to achieve trade-offs between efficiency and accuracy simultaneously,our method makes a novel combination of a segmentation stage(stage1),a detection stage(stage2)and a finetuning stage(stage3),and the first two stage are consisted of two fully convolutional networks(FCN)separately.In the segmentation stage we use a lightweight FCN to make a spatially dense pixel-wise prediction to inference the area of defect coarsely and quickly.Those predicted defect areas are the initialization of stage2,guiding the process of detection to refine the segmentation results.In the detection stage,we design a multi-loss-function structure to combine semantic information from a deep coarse layer with appearance information from a shallow layer.We also use an unusual training strategy.In the training phase,we use patches cropped from the training data.However,in the testing phase of stage1,we use the whole test images.We further show that the receptive field of the top layer has crucial influence on this training strategy.In the fine turning stage we use the guided filter to matting the area of defect.Besides that,we use depthwise separately convolution unit,stride convolution unit and upsample convolution unit to take the place of standard convolution layer,pooling layer and the deconvolution layer separately,aiming to reduce the cost of our algorithm.We will validate our findings by analyzing the performance we obtained on the dataset of DAGM 2007.We will validate the effectiveness,rationality and superiority of our algorithm.We also add noise in test date to validate the robustness.Our research makes a successful connection between the deep learning and the defect inspection,with the 512×512 input image,we can finish the segmentation task in 25 fps with the pixel accuracy above 99%.