Research On Real-Time Detection Technology Of Aluminum Profile Surface Defects Based On Deep Learning

Industrial aluminum is widely used in equipment manufacturing,building decoration and other life due to its light weight and good conductivity.There are many manufacturing processes of aluminum profiles,including production,molding,spraying,etc.,resulting in complex and diverse defects in the processing process,which easily affect the appearance experience and quality of products.Therefore,defect detection is an indispensable process in the production and manufacturing of aluminum profiles.Nowadays,the majority of manufacturing defect detection relies heavily on the quality inspectors’ experience,the conditions for artificial judgment of defects are not unique,and the detection performance is poor.This research focuses on the automatic surface defect detection method for aluminum profiles using computer vision technology and deep learning.Machine vision mainly detects defects through traditional image processing.This method has a lot of manual processing,poor versatility,and it is appropriate for straightforward single target detection tasks.This research suggests a multi-scale real-time detection approach for aluminum profile defects through the development and use of deep learning algorithm in order to address the issue of complicated aluminum profile surface defects.The following are the primary research findings:First of all,various methods of defect detection are analyzed and summarized.Aiming at the problems of multiple types of surface defects and large scale differences of aluminum profiles,the optimization strategy of deep learning model is designed,and the data set of aluminum profiles is expanded and divided.Considering the practicability and cost,this paper takes YOLOv5 algorithm as the basic model,and compresses,deploys and designs the algorithm.Secondly,in order to solve the problem that YOLOv5’s deep feature extraction network is easy to lose small targets and details,the CSPCA network is designed to consider the channel dimension and spatial dimension of coordinate attention in parallel,and embed the location information into the channel attention to reduce the interference of useless information;The adaptive feature fusion ASFF algorithm ensures the multi-scale feature invariance of the algorithm;Decoupled head is used to solve the problem of conflict between classification and regression tasks during YOLOv5 s detection;The EIo U function is optimized to accurately locate the defects of aluminum profiles.Finally,migration training and data enhancement methods are used to shorten the training time of the model.The performance of YOLOv5s-CASD model is verified by ablation experiments and different algorithm experiments.Considering the landing application of the algorithm,model compression and lightweight deployment are used to reduce the equipment cost,and the software system and images collection platform for defect detection are designed.In the aluminum profile surface defect detection task,the m AP of YOLOv5s-CASD algorithm is 96.4%,the speed of the defect detection is 84.1f/s,and the size is 7.2MB.The accuracy rate has increased by 5% when compared to the prior improvement,the speed is147.3%,and the model volume is 44.4%,which has better industrial application value.