Research On Defect Detection Algorithm Of Automobile Piston Based On Deep Learning

Since the beginning of the 21 st century,China’s auto industry has witnessed rapid development.The huge volume of auto production and sales has created a very broad market for the development of domestic auto parts industry.As the main part of the engine,the quality of the automobile piston directly affects the quality and performance of the whole automobile,and directly affects the life safety of the passengers.At present,the detection of automobile piston surface defects at home and abroad is mainly based on manual detection,which requires high detection accuracy and large workload.It leads to the decrease of production efficiency,the increase of personnel input and the increase of labor cost.In order to solve the above problems,this paper takes the engine piston as the research object and develops a deep learning method based piston surface defect detection algorithm for the surface defects produced in the production process.The research work and contributions of this paper are as follows:(1)Preparation of image data sets.First using automated image acquisition device,the piston surface defect image acquisition,followed by the use of Labelimg software defect data for tagging,and then select the possible defects in the real scene image amplification,and generate with original image expansion and change the lable information,which can greatly reduce the label such as mechanical work.(2)By analyzing the characteristics of piston defects,based on the detection accuracy and the advantage of deep neural network to extract the characteristics,the latest model complexity and deeper layer network was used to explore and research the piston defect.Then,the Resnet50 fusion network method was proposed and Efficient Net was applied.The results showed that the Efficient Net fusion model has a detection rate of 97.62% for cold insulation defect(sy)and94.66% for bump defect(jz),which were 0.65% and 6.87% higher than the non-improved model,respectively.(3)Due to the limited real-time performance of the deep and multi-repeated feature network,the application of the algorithm in industry is extremely difficult.Therefore,from the perspective of detection speed,the first-stage anchor-free YOLO V1,anchor-based YOLO V3 and SSD were used for the research.To test the performance of this kind of network in piston flaw detection,and make a good foundation for the research of lightweight detection network in the future.(4)In industrial applications,lightweight network helps to relieve the pressure on hardware resources and improve real-time performance.It is very important for industrial application to achieve a good balance between precision and speed of the algorithm.By comparing the deep neural network with the first-stage network with better real-time performance,the Faster R-CNN network is determined as the research focus of this paper and improved according to the needs of the project.It is mainly improved in four aspects,namely,replacing Backbone,finding the best Anchor parameters,replacing ROI Pooling with ROI Align and designing loss function.The experimental results show that the MAP of the improved network reaches 67.25%,14.81%higher than the original network,while the detection frame rate remains unchanged.Compared with other networks,the improved Faster R-CNN network achieves a good balance in terms of real-time performance under the premise of ensuring accuracy,and can be applied to industrial defect image detection.(5)In order to effectively test and improve the performance of Faster R-CNN network,the public defect data set was selected for experiments,and compared with the paper method using the public data set,the network performance of the Faster R-CNN network was obtained.The data set and public data set experiments of this project showed that the improved network met the requirements of industrial testing The algorithm can not only be applied to this topic,but also can be applied to other industrial image data set detection,with high robustness and generalization ability,which has a certain promotion and reference value for industrial image defect detection.