Algorithm For Defect Detection And Segmentation For Steel Based On Improved U-Net

At present,the domestic steel industry is facing the problem of overcapacity,and there is a strong demand for improving quality and efficiency.As an important part of ensuring product quality,surface defect detection still relies on manual inspection methods with low efficiency and strong subjectivity.With the continuous development of deep learning,the defect detection platform based on machine vision has begun to be applied on a large scale,which has improved the quality and production efficiency of industrial products.Intelligent development has become the theme of a new round of industrial innovation.In order to realize the automatic detection of typical defects on the steel surface,this paper improves the semantic segmentation network U-Net,designs an end-to-end defect detection segmentation algorithm,and improves,innovates and optimizes the algorithm from multi-scale feature extraction,segmentation network architecture,model compression and acceleration,and semi-supervised learning,etc.First of all,the characteristics of steel data set and task requirements are analyzed,and the whole scheme of defect segmentation algorithm is designed.The basic core modules including multi-stage data enhancement and preprocessing,feature extraction network,segmentation network and weighted loss function are analyzed in detail.Secondly,for the problem of large defect scale span and serious interference,a residual network architecture based on group convolution and feature channel shuffle is proposed to achieve good multi-scale feature extraction.At the same time,the problems of u-net are analyzed deeply,and through the improvement of transition features and dense connections,the inverted triangle fine defect segmentation network Nabla-Net is obtained.Attention mechanism is embedded in the improved network to adjust the feature channel and space adaptively to further improve the accuracy of the model.In this paper,based on the deep supervision mechanism,network pruning is performed on the NablaNet with too many model parameters,which effectively compresses the model and accelerates network convergence.Finally,in order to reduce the memory consumption of a single image during training when the hardware computing resources are limited,two numerical precisions of FP16 and FP32 are mixed in the training process to achieve the acceleration of training.In view of the difficulty of obtaining labeled data,this paper optimizes the algorithm with semisupervised learning based on pseudo-label technology,and uses the useless unlabeled data to improve the generalization ability of the model.In the experimental verification,a comparative experiment was designed on several network improvements and algorithm optimization in this paper,and the results were visualized and analyzed.The experimental results prove that the effect of the steel defect detection and segmentation algorithm proposed in this paper can meet the actual industrial production needs,with flexibility,practicality and effectiveness.