Improving Single Shot Detector for Industrial Cracks by Feature Resolution Analysis

Although the single shot detector (SSD) is effective for object detection in natural images, it is not suitable for special tasks such as the industrial crack detection. The difficulty lies in the wide diversity of crack sizes and shapes that is usually unpredictable. To solve this problem, we improve the SSD model by feature resolution analysis. The classical SSD network extracts several convolutional feature layers with degressive scales, and then classifies and locates targets by a series of prior boxes with default sizes and aspect ratios regarding to each scale. Therefore, the key is whether the design of these prior boxes is consistent with the real target characteristics. In this paper, we improve the architecture of SSD network via statistically analyzing the distribution of sizes and shapes from our collected crack samples. According to the resolution analysis of the targets at each feature scale, only a fewer number of valid feature layers are carefully extracted, and some more accurate prior boxes are designed relative to each scale. Finally, experimental results demonstrate that the proposed method could not only achieve significantly better prediction accuracy, but also acquire higher computational efficiency, which outperform the state-of-the-art methods.