Research On Industrial Thin Film Defect Detection Technology Based On Deep Learning

With the continuous development and application of new materials,industrial thin films have gradually gained widespread application in many fields.High quality film products have extremely high requirements for defect control,therefore,the detection of film defects has become a key link to ensure product quality.However,there are three major difficulties in defect detection of industrial thin films: firstly,due to the complex background interference,large differences in defect morphology,and varying sizes of defects in industrial thin films,the detection accuracy of traditional methods cannot meet the production needs of today’s enterprises.The second is to design a fast and accurate defect area calculation method based on twodimensional images,so that enterprises can analyze and adjust production in a timely manner.Thirdly,due to the large size of industrial films,the number of pixels captured in defect images has reached billions,requiring a method that can quickly process defect images while ensuring detection accuracy.This thesis has conducted detailed research on the above three difficulties,and the main work completed is as follows:(1)For defect recognition,this thesis proposes an improved industrial thin film defect recognition method based on Resnet50,which combines multi-scale deformation convolution and attention mechanism.Firstly,the convolution block of Bottle Neck structure in Resnet50 is refined into a group of multi-scale convolutions to expand the Receptive field,and then deformable convolution is introduced to replace the convolution kernel in the group,so that the network can capture the characteristics of flaws with different morphological scales in training.Finally,an attention enhancing module is introduced into the network,allowing it to focus on more important channels and spatial locations.Through a series of comparative experiments on a dataset composed of 1708 defect images collected in actual production,it is shown that the proposed method achieves an accuracy of 96.48% in industrial film defect recognition,which is superior to existing popular algorithms such as Efficientnet and Shufflenet v2.(2)This thesis designs a defect area calculation method based on two-dimensional images for defect area calculation.Calculate the true distance represented by a pixel by knowing the true distance and number of pixels of the standard reference object,and then combine the number of pixels contained in the defect area with the scale to calculate its true area.(3)This thesis develops a defect detection system on the Windows platform based on C #.In addition to the defect recognition,area calculation,and large image stitching methods,it also includes functional modules such as project management,database interaction,statistical analysis,camera control,and software registration.After enterprise acceptance and testing,it can be practically applied in enterprise production.