To address the problems of low detection accuracy of existing micro-motor armature surface defect detection methods,especially the misjudgment of similar workpiece,a new surface defect detection method for the armature based on improved BiFPN was proposed combined with deep learning method.The armature ROI image collected by the industrial camera was obtained by matching algorithm after cutting,the ROI was input into the EfficientNet structure,and the basic feature was extracted;the extracted features of different dimensions were fused by the improved BiFPN structure,and the features were screened;the classifier was used to output the final detection results.The results show that the armature surface defect detection method is more accurate than deep learning methods such as ResNet and EfficientNet,the accuracy is as high as 98.42%.The research results have positive significance to improve the detection performance of micro non-standard workpiece with great similarity.