Small Defect Detection Using Convolutional Neural Network Features and Random Forests

We address the problem of identifying small abnormalities in an imaged region, important in applications such as industrial inspection. The goal is to label the pixels corresponding to a defect with a minimum of false positives. A common approach is to run a sliding-window classifier over the image. Recent Fully Convolutional Networks (FCNs), such as U-Net, can be trained to identify pixels corresponding to abnormalities given a suitable training set. However in many application domains it is hard to collect large numbers of defect examples (by their nature they are rare). Although U-Net can work in this scenario, we show that better results can be obtained by replacing the final softmax layer of the network with a Random Forest (RF) using features sampled from the earlier network layers. We also demonstrate that rather than just thresholding the resulting probability image to identify defects it is better to compute Maximally Stable Extremal Regions (MSERs). We apply the approach to the challenging problem of identifying defects in radiographs of aerospace welds.