Detection Of Hub Identification And Valve Hole Positioning Based On Machine Vision

The rapid development of China’s automobile industry naturally led to the development of the hub industry.In the process of hub production,manufacturing and transportation management,it is necessary to identify hub type and locate its valve holes,the manual inspection is increasingly unable to meet the needs,intelligent automated hub manufacturing and detecting technology will further dominate the market.Machine vision involves numerous industries such as industrial manufacturing,medical image analysis and assisted driving.It is currently in a stage of rapid development and has great potential for development.Therefore,aiming at the problem of hub identification and valve hole positioning in industrial production with slow speed and low precision,a multi-scene hub identification and valve hole positioning detection system based on machine vision is designed.Firstly,the feature detection method based on image edge and texture features is studied,with emphasis on image smoothing,image sharpening,hub contour circle extraction,histogram of oriented gradient and local binary pattern feature extraction for hub type classification,spot feature extraction for valve hole with derivation and differential method and feature classifier.1000 sample images of 10 kinds of wheel patterns are randomly taken,and the hub recognition rate reach 99.3% by integrating histogram of oriented gradient and local binary pattern feature vector,and the valve hole positioning accuracy is up to 96.3%.A single piece is detected for about 700 ms on average,and the problem of hub identification and valve hole positioning is solved to a certain extent.At the same time,processing the abnormal scene of wheel hub image is more suitable than deep learning,but in standard industrial scenarios,there are still deficiencies in detection efficiency and accuracy.Secondly,the hub detection method based on deep learning is studied.The data set is constructed and enhanced by hub data preprocessing,and the feature extraction is carried out based on the improved Mobile Net V3 model and other lightweight convolutional neural networks.The single stage parallel strategy of classification and segmentation is adopted to directly obtain the hub category and valve hole location information through single detection.In the classification,the support vector machine is added for supervision,and the recognition of wheel hub type is completed.In the segmentation,the encoder/decoder structure is introduced to integrate multi-scale features,learn efficient classification,and complete the positioning of valve holes.By means of semi-supervised knowledge distillation,the model is compressed and accelerated,saving human resources by reducing the cost of labeling and perfecting the performance of the model at the same time.A high recognition rate of 99.9%F1-score for classification and 97.37% IOU score for segmentation was obtained on the test set.The detection of single hub image only needs about 27 ms on average.Finally,the real-time hub image detection system is developed according to the hub detection requirements of different application scenarios.The strengths and weaknesses of the two schemes are discussed and analyzed from the aspects of algorithm accuracy,speed,robustness and expansibility,and the deep learning detection method is determined as the scheme in the standard scenario.The edge texture feature detection method and deep learning detection method are integrated as the solution in non-standard scenarios,and the hardware and software design and hardware selection of the system are further determined.The model is deployed to the CPU device by using Open VINO tool,and the integration of hub identification and valve hole positioning detection technology is realized.