Abstract: The impeller blades of the engine are a key component of the propulsion system of a space spacecraft and play an important role in the success and efficiency of space missions. In order to solve the above problems, this paper proposes a defect detection algorithm (EF CenterNet) that integrates multi-scale features and attention mechanism, and uses the lightweight EPSANet network as the backbone of the CenterNet algorithm to effectively integrate the PSA segmentation attention mechanism, pay attention to more important defect features, and enhance the feature extraction ability of the network. At the same time, the FPN structure is added after the feature layer output by the backbone feature extraction network to further integrate multi scale information, that is, low resolution high level semantic information and high resolution low level feature information, so as to improve the defect detection accuracy of the algorithm. Experimental results show that the proposed EF CenterNet algorithm achieves an average accuracy of 96.74% in the self made dataset, which is 1.81% higher than that of the baseline CenterNet algorithm, and an average accuracy of 77.37% in the public dataset.

Key words: impeller blade, defect detection, attention mechanism, multi scale, deep learning