Hierarchical multi-scale network for cross-scale visual defect detection

Nowadays, an increasing number of researchers apply deep-learning-based object detection methods to implement visual defect detection in industrial manufacturing. However, large-scale variation in visual defect detection impedes the improvement of detection accuracy to be further explored. Therefore, we propose a hierarchical multi-scale block (HMS-Block), equipped with hierarchical representation and multi-scale embedding, to afford scale-abundant features to facilitate multi-scale defect detection. Specially, the hierarchical representation is implemented by a cascade learning stage to extract features from local to global at the channel level. Based on this representation, a cross-branch shortcut is concisely embedded to relieve the large-scale variation problem. Ultimately, the hierarchical multi-scale network (HMSNet) is published elegantly via stacking a certain amount of HMS-Blocks. The proposed methods facilitate the defect detection at all scales and outperform the ResNet50 baseline by a large margin with minor time overhead and less parameter required, indicating that the proposed HMS-Block has a high practical utility in the field of industrial applications. Moreover, the proposed HMSNet can also be applied to other detection-based tasks and greatly surpasses existing methods. Concretely, the proposed HMSNets achieve 42.4/42.7 mAP on NEU and COCO datasets, surpassing the recent backbones (i.e., HRNetV2) by 2.6/1.2 mAP.