A study of an improved RCNN network model for surface defect detection algorithm of precision workpiece and its realisation

Abstract

In this paper, an improved RCNN algorithm for surface defect detection of precision workpiece is proposed. During model training, the original image is processed by data amplification, which overcomes the phenomenon of model overfitting caused by too small data set. The fuzzy image is processed by the conditional generation antagonism network, which guarantees the authenticity of the image and eliminates the fuzziness of the image. On the basis of network selection, through network training and comparative analysis, ResNet101 was selected as the basic network for feature extraction. By combining the network performance, the number of convolutional layers (model depth), system efficiency and other factors, and taking advantage of the spatial characteristics of images, the irregular cross convolution kernel idea and the differentiated convolution kernel design method are adopted to realise the feature fusion of different convolutional layers. Embedded in the target Network Squeeze and Excitation (SE) module channel features fusion module, the Feature Pyramid Network (FPN) module, ROI Network module. The experimental results show that the improved network model is used to realise the location and classification of defects on the workpiece surface. The detection accuracy was 91% and the recall rate was 93%.