Research On Metal Surface Defect Detection Based On Improved Faster R-CNN

Nowadays,the surface defect of metal material products has attracted people’s attention,because the appearance of various defects not only affects the appearance and life of products,but also is more likely to cause serious quality problems leading to accidents,endangering the safety of people’s lives and property.There are many methods to detect the surface defects of metal materials,among which the traditional manual eye inspection is not only inefficient but also has a high rate of false detection and omission,so it is not suitable for mass production mode;The generalization of machine vision based on image processing is poor,and one model cannot identify all the complex and diverse defects in actual production.Therefore,the rapid and accurate identification of metal surface defects has important research significance.Based on the principle of deep learning and the idea of artificial intelligence,this paper studies the aluminum profile,which is most widely used in metals.Firstly,the status quo of the current detection algorithm is analyzed,and then the Faster R-CNN algorithm is studied.On this basis,three improvements to the algorithm are proposed,which are as follows:(1)For feature extraction with the deepening of network layer in the network,the characteristics of the small target information is missing,is proposed in this paper using the characteristics of the pyramid and multi-scale detection method instead of the original training top feature,so that the feature extraction can be extracted in the deep web to network more rich semantic information,reduce the number of small target omissions.(2)Due to the wide variety and large size difference of aluminum profile surface defects,the size of anchor frame generated in the candidate region generation network is not applicable to all defect types,resulting in inaccurate coordinates of the positioning frame of defects.In view of this problem,this paper proposed to use K-means algorithm to cluster the defect sizes in the data set,and five new anchor frame sizes were obtained,which improved the training speed and detection accuracy of the algorithm model.(3)In terms of training methods,this paper introduces the online difficult sample mining technology,which avoids the problems such as insufficient sample number,poor training effect and great training loss caused by subjective adjustment of positive and negative sample ratio,thus greatly improving the network performance.Experimental results show that the mean Average Precision(m AP)of the improved Faster R-CNN algorithm is 6.15% higher than that of the original algorithm,which proves the effectiveness of the improved method.Finally,the paper summarizes the work content of the whole paper,and puts forward the direction and content for further research.