Research And Application Of Deep Learning-based Automotive Parts Detection Algorithm

The automotive manufacturing industry is a crucial sector in the Chinese economy,playing a vital role in driving economic growth.However,this industry faces numerous challenges,such as improving efficiency,reducing costs,and enhancing automation and precision levels in the production process.In complex production environments,factors like unstable camera poses or lighting conditions lead to a decline in detection accuracy.Additionally,existing algorithms have high computational requirements,making realtime detection and recognition difficult on devices with limited computing power,such as CPUs.This significantly affects the efficiency of industrial machine production processes.This paper explores the potential of computer vision in the automotive manufacturing industry,with the following specific research objectives:(1)Research on automotive parts detection algorithms for industrial production.To address the problem of reduced detection accuracy caused by camera pose changes,uneven lighting conditions,and the inability of traditional detection methods to balance recognition accuracy and real-time detection,this paper proposes an enhanced receptive field lightweight network.First,the Shuffle Netv2 network is used as the network backbone to balance accuracy and speed by controlling the width factor to reduce the network model.Second,an improved parallel atrous pyramid lightweight model is introduced to expand the receptive field and enhance the network’s feature extraction capabilities.Finally,the GSCSP module is designed in the path aggregation network to promote feature information transmission and enhance network fusion capabilities.In the industrial automotive engine compartment detection task,the improved algorithm’s m AP reaches 96.1%,and the network model training weight is only 4.3M,which improves by 24 frames/s compared to the YOLOv5 network.(2)Designing system software for automotive parts detection in an industrial environment.First,analyze the functional requirements of the visual detection system for industrial production of parts and determine the non-functional requirements such as real-time and accuracy that the system needs to meet.Then,perform a detailed design of the system’s functionality,including the camera configuration module,intermediate communication module,visual detection module,and system configuration module.Finally,use Python language and QT software to implement platform design,generate the improved lightweight network model as the detection core algorithm,and deploy it to the system as a dynamic library through industrial cameras reading collected images and inputting them into the improved detection model for inference.(3)Testing the automotive parts detection system for industrial production environments.Using actual automotive engine compartment datasets collected from industrial environments,test the system’s functional and non-functional requirements.First,perform comprehensive testing on the four functional modules of the system with test cases and compare the improved algorithm’s detection effectiveness with other algorithms.Additionally,test the non-functional requirements such as startup time,feature detection time,image saving time,and result rendering time.The test results show that visual detection is below 1500 ms,meeting the real-time detection requirements of production,and the component detection success rate reaches 99.5%.This system has improved industrial production efficiency to some extent.