An Accurate and Explainable Residuum Inspection Network With Residual Block Connection and Hybrid Attention Fusion

The residuum inspection is inevitable since it directly affects the quality of the product. However, it is a challenge to accurately inspect the residuum due to its complicated distribution and inhomogeneity of the particle size. To address these issues, a lightweight residuum inspection model is designed, which possess strong classification ability along with intuitive explanation performance. In specific, residual block connection (RBC) is designed in feature extraction network, which shows great promise in incorporating multilevel valuable features and preventing leakage of nonadjacent features. Besides, a hybrid attention fusion (AF) module combining channel attention (CA) and spatial attention (SA) is designed, which efficiently highlights useful features and weakens irrelevant features. Moreover, the output of the inspection model consists of classification primary head and explanation auxiliary head. The classification primary head introduces classification enhance module (CEM) and shares weights with the explanation auxiliary head, which makes the inspection results more precise and explainable. Finally, numerous test experiments are conducted in our established residuum inspection dataset (Res-DET) and DAGM dataset, and the classification accuracy reaches 99.9% with detection speed 159.67 fps. It indicates that the accuracy and efficiency fulfill the requirements in industrial production. Ablation and comparative experiments verify the necessity of each component in the established network as well the superiority on residuum inspection. Soon, we will make the source code and dataset Res-DET available online for public use.