Detection of Surface Defect on Impeller Blade Images based on Improved Centernet Algorithm

To solve the problems of low precision, poor algorithm robustness and high leakage rate of traditional image processing algorithms in the detection of impeller blade surface defects, and according to the characteristics of impeller blade surface defect, this paper designs an improved Centernet-based algorithm for impeller blade surface defect detection. First, the performance of 3 networks for feature extraction was compared and ResN et50 was selected as the backbone network; Then, the introduction of the attention mechanism (CBAM), which places attention on important feature information to extract more useful features without deepening the network; Finally, based on the constructed impeller blade surface defect dataset, and in order to better compare with other models and verify the generalization of the algorithm, NEU-DET, a public dataset, which is closer to the actual part data, was selected as the validation object, both use the improved Centernet algorithm for detection. The experimental results show that the improved Centernet algorithm reaches a mean average precision of 96.8% on the test set of impeller blades, which is 2.5% higher than the algorithm before the improvement, and can effectively detect a variety of typical defects. The mean average precision of the public dataset reaches 75.8%, the algorithm can meet the accuracy requirements for automated detection of different types of defects on the impeller blade surface.