Research On Surface Defect Detection Of Automobile Air Conditioner Liquid Reservoir Based On Deep Learning

Metal materials,as the main constituent materials of industrial components,are an important part of the industrial manufacturing process.Many types of uncontrollable defects occur on the surface of metal materials during processing,and their surface defects have an adverse effect on the quality and performance of industrial products.In order to meet the increasing demand,in the process of industrial production,it is necessary to strictly guarantee product quality while improving production efficiency.Therefore,the detection of surface defects of metal materials has become an important link in actual industrial production.At this stage,there are some detection methods for different types of defects on the surface of metal materials in industrial production.However,in the actual production process,the environment is more complicated,and factors such as light,temperature,and defect distribution all affect the detection effect.Manual inspection is still used in most factories.The detection method.However,manual testing relies too much on the experience of the testing personnel,and the efficiency of testing is low.A large number of samples cannot be carefully screened,and testing requirements in the production process cannot be well met.Therefore,research and development of a method for automatic detection and classification of metal material surface defects is of great significance.Through the understanding of computer vision,image processing,deep learning methods and traditional visual inspection and other related knowledge,this research designed a metal material that can meet the needs in terms of accuracy,efficiency and robustness based on deep convolutional neural networks.Automatic detection and classification of surface defects.In the specific research process,the open source NEU surface defect detection challenge data set was used as the benchmark data set to train common target detection models such as YOLO,SSD,FCOS and FasterR-CNN.By comparing the training results,the four models found that the highest accuracy detection results are from the FasterR-CNN model,with an accuracy of 90.4and an m AP of 84.9;and the most efficient model is YOLO v3,which requires a training time of 16 h,prediction The time is 0.96 s,the lowest Fps in the target detection process is 49,and the highest Fps is 95.Among the statistical results of AP value of 6 different types of defects detection,including cracks,inclusions,plaques,pits,scratches and rolling,the highest value comes from the SSD model in the detection of plaque type defects,the AP value reached 92.2;the lowest;The value comes from the FCOS model’s detection of rolling type defects,and the AP value is only 59.5.Further research decided to select the FasterR-CNN model with the highest accuracy for further optimization research.The FasterR-CNN model was optimized by using different feature extraction networks,adjusting activation functions,and referencing transfer learning methods.The improved FasterR-CNN model has a target detection accuracy of 0.9194 and a classification accuracy of 0.8693 on the NEU surface defect detection data set.The improved FasterR-CNN model is used to predict the images of air-conditioning reservoir parts collected at the normal industrial production site.The prediction results show that the improved FasterR-CNN model in this study is ideal for the six types of defects in the training set.The detection and classification results prove the robustness and feasibility of this research method.The proposed method is expected to provide key technical support and practical tools for the automatic detection and classification of metal material surface defects in real production scenarios.