Multi-label Instance Segmentation Network For Industrial Product Defects Detection With Complex Background

The technique of surface defect segmentation aboutindustrial products refers to the defects are distinguished by several segmentation algorithm automatically.It is indispensable for the production process to improve the quality of industrial products effectively.Currently,several segmentation methods for surface defecte segmentatio n can recognize some types of defects,but it still exists some challenges as follow.Firstly,some industrial products with metal materials appear different degrees of light reflection,due to the interference of a variety of complex light sources Specifcially,current detection algorithm is also easy to misjudge this kind of reflection characteristics as defect characteristics.Secondly,different from common products,the surface of industrial products contained with rich texture and colorful features,such as cartoon texture.Current algorithms with small object detection can distiguhish some small scale defects,but it is too sensitive to edge features,resulting in some subtle scratches with unclear defect features.Therefore,in order to solve the situations mentioned above,this thesis proposes an image surface reflection suppression algorithm and a new example segmentation network to improve the above problems.The specific algorithm is described as follows :（1）In this thesis,we proposed an adaptive light source algorithm based on spectral distribution for image surface de-reflection.Firstly,the image surface texture features are enhanced by wavelet transform,then the reflective position of the image is located by gradient recognition.Finally the light source of the image is adjus ted adaptively,so as to suppress the interference of the reflective area to the image surface defect information.（2）In this thesis,Rep-Segnet network model is proposed to detect complex surface defects of industrial products.Firstly,in order to solve the problem of insufficient feature extraction ability of the network in the case of the complex surface texture of industrial products,Rep-segnet adopts Rep-Encoder encoder and uses multibranch residual structure to enhance the feature extraction abili ty of the network model.Secondly,in order to overcome the problem that it is difficult to detect narrow,long and small product surface defects,multi-scale spatial attention mechanism（MSAM）is adopted to make full use of the spatial information between different scales,strengthen the spatial information of complex pattern surface defects,and improve the detection accuracy of narrow,long and small defects.In this thesis,ablation experiments and comparative experiments are carried out on the network proposed in the complex surface defect data set of industrial products.The experimental results show that the adaptive light source algorithm based on spectral distribution can effectively suppress the reflection phenomenon of image surface.Rep-Segnet can accurately segment the defect target and accurately identify the defect category.