Simulation of Industrial Product Surface Defect Detection Based on Faster R-CNN

The traditional surface defect detection technology for workpieces has problems such as missing small target defects and low detection accuracy. Therefore, an improved object detection model based on fast region convolutional networks is constructed. This model innovatively constructs a multi-scale feature extraction network to integrate features from different levels, introduces an iterative clustering analysis K-Means algorithm to reselect anchor box sizes, and uses Online Hard Example Minging (OHEM) technology to retrain candidate regions. The OHEM technology utilizes a selection process that involves the most challenging samples for gradient updates, thereby providing an assessment of the model’s overall performance. To verify the effectiveness of the strategy, a simulation experiment of surface defects on the workpiece was designed, and the mean average precision (mAP), which can comprehensively measure the overall performance of the model, was used as the evaluation index. The results showed that the accuracy, Missrate, and False Alarm of the improved model on the hot-rolled strip dataset were 93%, 3.9%, and 2.5%, respectively. Its average precision was close to 95%, with a mAP of 93.5% for various types. These three values on the diode glass shell dataset were 97.9%, 3.4%, and 1.9%, with a maximum average precision of 99% and an mAP of 98.9% for each category. The three values on the surface defect dataset of titanium alloy pressed parts were 92.5%, 3.4%, and 1.9%. The highest average precision was close to 95% and the mAP of various types was 92%. Research has shown that the multi-scale object detection accuracy of this model has been greatly improved, which can more effectively carry out the inspection and detection of surface defects on workpieces. The research results will help improve the recognition ability of surface defects on complex workpieces and meet the rapid detection needs in industrial production.