Research On Image Segmentation Algorithm Of Metal Surface Defects Based On Visual Attention Mechanism

Metals are of vital importance as the raw materials and products of current manufacturing and industrial process.Its performance would affect the following products in line with product quality,industrial chain,and even public safety.Therefore,the automatic detection of deficits on metal surface makes an significant effects to the industrial process of metal quality control.Current technology of detection applies the machine version method,having become the main-stream in industrial use.The machine version is originally imitating the human visual system with hardware equipment and underlying algorithm.The underlying algorithms of image pro-cessing for machine version detection are now focused as the mature detection equipment are developed.The good performance of underlying algorithm will become the key factor affecting the deficit detection system.Considering the weak contrast of deficit region to the background,various shape of deficits and interference of highlights and shadows in the images,current algorithms present significant rate of detection error.To improve the current application of machine version in industrial pro-cess,this work will present an investigation on the development of new segmentation algorithm for metal deficits in surface images,and the validation of algorithms for deficit detection.The segmentation algorithm simulates the mechanism of human attention by machine version,with considered the calculation cost and segmentation efficiency.The investigation comes out some concluding re-marks as follows:(1)In view of the lack of metal surface defect data set,a data collection and production plan for copper strip surface defects was designed,the data set was marked,and the copper strip surface defect data set(Kmust-DET)was established.Then extracted four main image characteristics,using three types of feature Bayesian classifier classified by the classification results can be found in the design data set image quality and suitable as a research goal.(2)A traditional segmentation algorithm based on visual attention mechanism is developed to extract the image features.Among which,the improved Retinex algorithm is applied to correct the gray level of the metal surface defect image,and the saliency mapping is formed with the classified global and local information.With combining the saliency mappings,the image segmentation is processed relying on the critical thresholds.The robustness and efficiency of this algorithm is confirmed with experimental methods.(3)A semantic segmentation method for metal surface image with deficits is proposed.This work investigated the attention mechanism with image features and existing semantic segmen-tation architectures.With applied various attention modules in different stage of image process-ing,this work improves the central network of algorithm and develops a new branch network.The stability of this method is confirmed accordingly with segmentation results,Accuracy rate of 91.34%.