Research On The Method Of Aluminum Profile Surface Defect Detection Based On Deep Learning

As a common and important metal material,aluminum profiles are widely used in national defense,aerospace,construction,automobile,electric power and other fields.However,during the production process,some aluminum profiles inevitably produce defects on the surface,which affect the product quality and safety.Therefore,it is very necessary and challenging to detect defective aluminum products on the production line in a timely and effective manner.Traditional inspection methods have many limitations.For example,manual visual methods are inefficient and subjective,while image processing methods require extensive a priori knowledge and parameter adjustment,making it difficult to adapt to different types and degrees of defects.In recent years,researchers have made breakthroughs by applying deep learning methods to the field of computer vision.At the same time,deep learning methods have been introduced to detect surface defects in aluminum profiles.However,in the industrial production environment,the real-time detection speed is increasingly demanding,and because the defects of aluminum profiles have problems such as small and inconspicuous targets,they need detection models with complex network structure and high accuracy,which is in contradiction with the limited computing power and storage capacity of industrial field devices.To address these problems,this thesis applies deep learning technology to the task of detecting surface defects in aluminum profiles,and focuses on several directions:model lightweighting based on backbone network replacement,model optimization based on enhanced feature extraction network improvement,and model optimization based on loss function,so that the model meets the requirements of the defect detection task,and develops a corresponding detection tool using the front-end and back-end frameworks.The specific research of this thesis is as follows:(1)Selection and lightweight study of detection models.In this thesis,we first analyze and consider the application of target detection models in aluminum surface defect detection tasks,select three representative and high-precision target detection models as the candidate models for aluminum defect detection research,and select the YOLOX detection model with the best comprehensive performance according to the experimental results.Secondly,to address the problems of the current detection models in terms of large computation,large number of parameters and low real-time performance in aluminum surface defect detection tasks,this thesis proposes a lightweight improvement scheme under the premise of ensuring the accuracy of the YOLOX model as much as possible,using the intercepted Mobile Net series lightweight network as the backbone network of the YOLOX model to improve the detection speed of the model and reduce model resource consumption.The experimental results show that the obtained lightweight detection model has the best results after incorporating the Mobile Netv2 network.(2)Model recognition effect research.To address the problem that small target defects are difficult to identify and locate effectively,this thesis improves the enhanced feature extraction network by incorporating different lightweight attention mechanisms into the FPN+PAN structure of the enhanced feature extraction network,respectively,to enhance the detection model’s attention to small target defect categories such as aluminum surface scuffs,dirty spots,spray flows and paint bubbles,to improve its recognition accuracy while ensuring the model size and detection speed and the ECANet attention mechanism is selected as the best choice by experimental comparison.(3)Model loss function optimization.For the problem that the model has too many negative samples in the training process,which makes the model training difficult,this thesis introduces the Focal Loss loss function in the model detection head to improve the model’s learning of key defect features in the training process,reduce the influence of background negative samples on the model training,and increase the training weights of difficult classification samples,so that the model focuses on the samples that are easy to be misdetected in the learning and improve the model training effect.In order to verify the effectiveness of the improved model proposed in this thesis in the task of detecting surface defects in aluminum profiles,this thesis conducts a crosssectional comparison experiment between it and the YOLOv5 m model with similar volume on the same data set.The experimental results show that the detection model proposed in this thesis improves the average accuracy(m AP)index by 5.16% compared with the YOLOv5 m detection model with a slight lead in detection speed,model computation and model parameter number,which verifies the effectiveness of the improved model proposed in this thesis.