Research On Tire Damage Identification Based On Computer Vision Technology And Deep Learning Algorithm

Tires are one of the key components that directly support the ground and affect the stability of vehicle driving in the process of vehicle driving.Due to road conditions and driving habits and other factors,different types of damage will inevitably occur during service.When damage occurs,vehicle owners usually cannot timely obtain information on the health status of the tire,which poses a serious threat to road traffic safety.Therefore,in recent years,research on tire damage identification and warning for motor vehicles has become increasingly important.Most traditional tire damage warning methods rely on high-precision sensors,but there are problems of high cost and inability to apply to different types of tires.Existing image-based methods cannot effectively identify tire surface interference information and weak representation,etc.Therefore,based on computer vision technology and deep learning algorithms,this article proposes and developed a real-time tire damage recognition system for vehicles in low-speed driving conditions.The main research content is as follows.First,this paper provides the establishment of damage tire image database based on the relevant database of cooperative tire enterprises,and investigates and counts the corresponding causes,mechanisms and characteristics of common damages.Based on it,this paper classifies tire damage into three major categories,which are bulge,crack,and fragment embedding.The texture features of damage images are extracted,the correlation law between each damage texture feature and category is analyzed,and the influence of each texture and image enhancement technique on the correlation between texture and sample category is explored.Second,a tire localization detection algorithm based on improved circular Hough transform and perceptual hashing is established.For the problem that the original algorithm cannot exclude concentric and non-concentric interference circles,a tire typical feature library is established,and tire accurate localization detection of multiple structures and uses is realized based on similarity calculation of multiple suspected tread areas.Tire detection frame and interference frame data sets are established and experiments are designed to verify them.The results show that the tire localization detection algorithm locates better than the traditional single feature matching algorithm,has strong anti-interference capability,can operate under a variety of conditions such as inconspicuous tread color,low resolution,strong and weak lighting,and effectively excludes the interference of irrelevant information such as road tread pattern and vehicle structure.Thirdly,an efficient tread damage recognition network CA-Eff Net(Composition Augmentationt-Efficient Net)with Efficient Net as the main trunk network is proposed to solve the problem that common models cannot effectively recognize the damage with weak representation,darkness and weakness,attachment interference and so on.The network uses strong feature transformation and contrast fusion to achieve repeated learning of the same image under different feature transformations to increase the damage features obtained by the network.Contrast fusion of the original features eliminates the performance damage caused by random parameters and excessive feature enhancement,and optimizes the feature mapping selection of the network.The ”backtracking-weight freezing” method is used during the training process to assist in the parameter search of the feature enhancement structure and to eliminate the harmful feature transformations.The CA-Eff Net can effectively remove the influence of tread color,wear,stain adhesion and other factors on damage recognition,and improve the recognition ability of weakly characterized damage such as initial bulging and small cracks significantly.The overall feature extraction ability is better than that of the pre-improvement network.Finally,an in-service tire damage identification system was built.The system is composed of industrial-grade Gig E camera,ground magnet,and fill light,etc.,which can realize clear imaging of tires at 16KM/h under 25 FPS continuous acquisition.The system first uses CHT-PH tire location detection algorithm to locate the tire area in the image,and then uses CA-Eff Net efficient tread damage identification network to identify the tire tread damage and real-time warning.The results show that the system effectively excludes the interference of vehicle structure and background on tire positioning,and strips the influence of tread color,wear and stain adhesion on damage recognition.The average time for processing a single image with this damage recognition system is 0.7 seconds,which has the advantages of high recognition accuracy and fast processing speed,and is suitable for large-scale deployment in scenarios with low-speed vehicles such as toll stations and gates.