Research On High Resolution Defect Detection Method Of Metal Surface Based On Convolutional Model Selfo-Ptimization

Metal is a kind of widely used industrial manufacturing material,which is easy to produce surface defects in the process of production and processing.These defects have a great impact on the quality of metal products.Therefore,accurate detection of metal surface defects is of great significance to ensure the quality of metal surface.With the development of artificial intelligence and deep learning,it has become an inevitable trend that machine automatic detection replaces manual detection.At present,deep learn-based metal surface detection still has the following problems to be solved:(1)The resolution of defect images collected from metal industry production lines is low,and the low resolution defect images affect the accuracy of defect detection.How to enhance the edge and texture details of metal surface defect images is the key problem to improve the detection accuracy;(2)There are many small target and dense target defects in metal surface defect images.It is also a technical difficulty to design a high-performance detection model to detect small target and dense target when considering model calculation cost and a small number of defect data sets.(3)There are a large number of super parameters in the super resolution reconstruction model and the defect detection model,which affect the performance of the model to a large extent.At present,most of these super parameters are artificially adjusted by researchers and lack of model adaptability.How to achieve adaptive adjustment of model super parameters is also the focus of this paper.To solve the above problems,this paper carries out the following research contents:(1)Research on image high perception super resolution reconstruction method based on multi-layer fusion model.With ESRGAN as the main stem network,a refined layered structure was constructed to enrich deep features,and an edge enhancement module was constructed to enhance image edge features.Meanwhile,a loss function based on edge enhancement module was constructed to reduce reconstruction errors.Experiments show that the PSNR of the superresolution reconstruction model constructed in this paper can reach 30.932 on the NEU-DET test set,and has good generalization ability after testing on the common data set.(2)Research on metal surface defect detection algorithm based on bottleneck residual dense structure.Based on the YOLOv5 model,a bottleneck residualintensive structure was constructed to enrich the multi-scale features of the backbone feature extraction network and realize feature reuse,thus improving the detection accuracy of the model for small targets and dense targets.Experiments show that the detection model m AP0.5 of metal surface defects can reach 0.782 and the detection speed can reach 102 fps.(3)Research on high-performance convolution model self-optimization algorithm.A multi-mechanism fusion algorithm was constructed to perform adaptive optimization of RRDB block number and generator loss function weight in the super-resolution reconstruction model,and Bayesian optimization algorithm was used to perform adaptive optimization of key super parameters of the metal surface defect detection model,such as the initial learning rate,momentum,loss weight of prediction frame,etc.Experiments show that the multimechanism fusion algorithm improves the PSNR optimized by multi-layer fusion model by 0.7-1.2d B on each test dataset.The Bayesian fast optimization algorithm improved m AP0.5 by 4.58% and MAP0.5-0.95 by 10.34% after the optimized detection model.