When using visual image technology to detect surface defects of sealing rings for aerospace, traditional imageprocessing algorithms are constrained by surface defects types, scale differences, illumination variations, andnon-uniformities of surface image. As a result, the application of each algorithm is relatively limited. To solve theproblems, a general deep learning algorithm for surface defects detection of sealing rings is proposed. Texture features ofdefects, gradient features of edges and contour features are selected as main features of surface defects of sealing ringsdifferentiated from the background, and a new cascade modular backbone network is design for extraction of thesefeatures. The average aspect ratio of label box for defects among dataset of defect images of sealing rings is calculatedand aspect ratio distribution is counted to obtain the proportional parameter for generating prior box. Based on thecalculated proportional parameters of prior box and scale features of defects, multi-scale defects detection network isdesigned. Meanwhile, defects classification and position regression compound loss function are defined. In the end,weighting parameters of the network are updated on the basis of end-to-end training. Experimental result shows that theaccuracy of proposed algorithm for defects detection reach 94.03% under the condition that the original defects sampleimages is 107 and confidence is 0.6. Compared with RefineDet network, the depth of backbone network of featureextraction increases by 78% while the scale of parameters reduces by 84.82% and the accuracy of defects detectionreduces by 3.01%.
展开▼
