Surface Defect Segmentation with Multi-column Patch-Wise U-net

Surface defect segmentation of the production plays an important role on the performance in industry. However, training the model by deep learning methods on industrial dataset with high resolution and few images is computationally impossible. Therefore, in this situation, we train a patch-wise fully convolutional neural network to perform similar to related image-wise methods. On this object segmentation task, we propose a method in inferring object segmentation by leveraging only ellipses annotation roughly indicating the defective area. This problem is viewed as a kind of weakly-supervised segmentation task, it can be solved by the approach that combines deep convolutional neural networks with the Multiple Instance Learning (MIL) framework. In view of this, we add the global pooling layer that called MIL-layer behind the last convolutional layer to compute the average class probability of all pixels in every patch. Finally, we combine three of this fully convolutional neural network with MIL-layer to become MIL-layer, which can utilize filters with receptive fields of different sizes. To address defect/non-defect class imbalance problem, we weight the loss contribution of false negative and false positive examples by our Weighted SoftMax loss function. The proposed network is evaluated on the detection of industrial inspection in the dataset of DAGM2007. Compared with some typical fully-supervised segmentation methods, our method achieves competitive accuracy (99.2%) but only requires easily ellipse annotations.