Algorithm Design And Implementation Of Industrial Appearance Defect Detection Platform Based On Stereoscopic Light Source

Nowadays,the surface defect detection of workpieces in many industrial scenes still needs to be carried out by manual means,it is difficult to use automated equipment to complete.This area remains a challenge for multiple industries.Since the largescale development of wide learning in 2013 in computing power and theory,there is hope for the resolution of such problems.Based on the four actual industrial defect detection scenes of battery surface,aluminum plate surface,automotive parts,and cloth,this paper makes many improvements based on the two-stage target detection algorithm and integrates the deep learning method into the industrial scene.The main research work in this article is summarized below:1.Through the light source experiment of the workpiece,the dynamic stereo light source is designed,which solves the problem that many defects cannot be imaged under a single light source.The use of multi-plot overlay joint detection means that each type of defect can be under the camera.2.To detect small targets and multi-scale targets as accurately as possible,three methods are adopted in this paper on the extraction method of multi-scale features.First,by using an image pyramid method before the image input network,the defect,regardless of size,is reduced to a suitable scale,making the detection network relatively easy to process,and second,the characteristic pyramid is used in the characteristic extraction method,respectively,in the detection of the main network and candidate box extraction network into the idea of feature pyramid,So that each layer feature map can contain multi-scale information,and then a correlation constraint is used in the last layer of the main network feature map,so that the nonrelevance of the feature map is enhanced,and more information can be characterized in the one-layer feature map.3.To make the improved algorithm in this paper better adapt to the industrial defect detection scene,the algorithm is improved for certain requirements.the main network improvement of multi-channel light source is the fusion of multichannel images can be combined training,and secondly for industrial detection pictures are mostly high-resolution pictures,the large footprint of the memory leads to a reduction in batch and thus affect the model distribution,the use of group regularization method so that the model can still be normal training in small batch cases.Finally,because of the need to extract the defect shape,the mask branch is added,so that the defect can be split to obtain the shape profile.4.Four data sets are constructed,and the improved algorithm can be wellperformed on all four data sets through experiments.Besides,a set of on-line detection system,a training system,a factory data management system and a set of semi-automatic labeling platform have been developed in engineering to help the factory to land the algorithm.Form a closed-loop solution.