Leather Defect Detection Based On Photometric Stereo And Image Saliency

With the rapid development of technology and increasing improvement of living standard,quantity demanded of leather products is on the rise.Thus,the quality monitoring of leather products become more and more important.Many defects will inevitably occur on the surfaces of leather raw materials during the production process,such as scratching,folds,loose grain,holes and color aberration.If these defects can be inspected effectively,the quality of leather raw materials can be ensured,which will result in a significant effect on the consequent production process.Nowadays,these defects are usually inspected by human vision.However,it will result in the problems of low efficiency and high error detection rate due to the disadvantages of subjective factors,labor intensity,low consistency and being influenced easily.In the recent years,with the rapid development of artificial intelligence,machine vision is successfully applied to defects detection in the real assembly lines.In order to improve the efficiency and automation of leather defect inspection,the inspection methods for leather defects are proposed based on photometric stereo and salient object detection.Also,an inspection system for leather raw materials is designed to solve the problems of low contrast of leather images with complex textures.In the system,appropriate image processing methods are employed,including image preprocessing,image enhancement,salient object detection,adaptive threshold segmentation,and leather defect detection and location.First,an image acquisition platform for photometric stereo is established with a scheme of flexible light source to acquire the leather images from different illumination angles.Second,surface normal vectors of multiple leather images are achieved by photometric stereo and then image reconstruction is implemented.Next,reconstructed surface normal vector image is filtered by a curvature filter.The curvature filter is considered as image enhancement since the curvature is a measure of geometric irregularities.However,for some defects such as color aberration,the original reconstructed image is more suitable for consequent image processing.Here,approximate surface roughness is used to measure the foregrounded information in the surface normal vector image.The surface normal vector image with more information is selected in the following image processing.Finally,saliency object detection based on spectral residual and adaptive threshold segmentation based on expectation of gray image are employed to implement the leather defect detection and location.The result shows that the leather defect inspection system can efficiently and accurately inspect the defects on the leathers' surfaces.Compared with the present leather defect detection method,the proposed inspection method can inspect many defects of various materials and texture more efficiently and accurately.