Study On Pseudo Defect Removal Method For Hot-Rolled Steel Strip Surfaces

Hot-rolled steel strips are important basic materials in the field of iron and steel metallurgy.Automated visual inspection(AVI)plays an essential role in ensuring the surface quality of hot-rolled steel strips.The core task of AVI is to realize the rapid identification and accurate localization of surface defects.However,AVI equipment is usually installed at the end of the production line,behind the laminar flow cooling tower and before the crimping machine,where exists many kinds of interference,such as waterdrop dispersion,waterline splashing,water mist,frequent vibration and so on.These " pseudo defects " are mixed with the real defects and induce a large number of false alarms,which seriously affect the detection performance of AVI equipment.Therefore,how to remove the surface pseudo defects of hot-rolled steel strips in complex industrial environment is of great significance to improve the performance of AVI system.This paper focuses on the research on pseudo-defect removal,and the specific contents are listed as follows:(1)In this paper,it is analyzed that the waterdrops,splashing waterlines and diffusing water mist on the high-speed rolled steel strip surfaces have "rain-like characteristics".According to different distribution properties,pseudo defects are classified into narrow pseudo defects and generalized pseudo defects.Various waterdrops scattered on the strip surface are defined as narrow pseudo defects and named as "droplet pseudo defects".However,the slender waterlines and white tiny waterdrops splashing in the imaging space above the horizontal plane of steel strips can also trigger false alarms,which are further included in the category of pseudo defects.The extended pseudo defects are named as "rain-like layer pseudo defects",which can better describe the actual pseudo defect characteristics of the production line.(2)Progressive recurrent generative adversarial network(PRe GAN)was proposed to achieve robust removal of "droplet pseudo defects".The network consists of two parts: a generator based on progressive recurrent network and a Markov discriminator,which achieve fine-grained image restoration by accurate tracking and location of dynamic waterdrops.PRe GAN realizes 52.2073 peak signal-to-noise ratio(PSNR)and 0.9502 structural similarity index(SSIM)respectively on the "droplet pseudo defects" dataset constructed in this paper.The experimental results demonstrate that PRe GAN is competent accurate identification of "droplet pseudo defects" and fine repair of missing information.(3)Attentive dual residual generative adversarial network(ADRGAN)was proposed to solve the problem of robust removal of rain-like layer pseudo defects.The network consists of three parts: a generator based on encoder-decoder structure,a guide based on mask image to enhance attention and a discriminator based on multi-scale features.It can focus attention on the part of pseudo defects,while retaining the edge and texture details of the attention-focused area,achieving fine repairing of the steel texture obscured by pseudo defects.The image restored by ADRGAN is the closest to the real images collected by the actual production line,with PSNR and SSIM reaching56.8627 and 0.9980 respectively.With the aid of ADRGAN,false alarms can be reduced by more than half by traditional defect detection methods such as threshold decision.