| As a basic material in the field of industrial manufacturing,the surface quality of aluminum profiles will directly affect the reliability and service life of industrial products in the fields of rail transit,machinery manufacturing,aerospace and aviation.However,due to the complexity of the production environment and the limitations of processing equipment,defects will inevitably occur on the surface of aluminum profiles.To ensure the long-term use of aluminum profile related products,timely detection of its surface defects is of great significance.On the one hand,it helps to extend the service life of the product.On the other hand,it will be helpful to find weak links in production and reduce the major economic losses caused by aluminum profiles surface defects(APSD).At present,the traditional detection methods for surface defects of aluminum profiles cannot be adapted to the quality inspection requirements in a complex production environment.Based on this,this paper aims at the APSD detection and conducts research on several aspects such as sample imbalance,noise interference and model lightweighting.The main research contents of the thesis are as follows:(1)Considering that the imbalance of common and rare defects on the surface of aluminum profiles increases the difficulty in identifying surface defects,a APSD detection method based on weighted deep residual focal network(WDRFN)is proposed.First,the weighted random sampling method is used to give a higher sampling weight to the minority samples and increase the training frequency of the minority samples.Then,the focus loss function is used to adjust the balance factor and focus parameters to improve the confidence of the output features of the minority class,and fully train the minority class samples.The experiment uses F-score as the evaluation index,and the experimental results reflect the reliability of the WDRFN.(2)Considering that it is difficult to detect APSD under mixed noise conditions,an aluminum profile surface defect detection method based on deep self-attention network(DSAN)is proposed.Firstly,the residual learning strategy is used to reduce the feature noise information of aluminum profile surface defects;then,the self-attention mechanism is used to add a corresponding weight matrix to the defect feature map to enhance the global correlation of the features.This mechanism further improves the feature extraction ability of aluminum profile surface defects under noise conditions.The experimental results show that the DSAN has good noise robustness under both single noise and mixed noise conditions.(3)Considering the large amount of parameters and the long training time of the APSD detection network,a deep lightweight attention network(DLAN)for APSD detection is proposed.Firstly,it uses small convolution kernels instead of large convolution kernels and reduces the number of input channels,reducing the amount of training parameters;Secondly,it adds a lightweight attention mechanism to the network output features.The mechanism uses cross-dimensional interaction to directly obtain channel and spatial feature weights to avoid increasing network training time;Then,weight learning is performed on output features to achieve refined extraction of defect features,improving feature extraction capabilities.The experimental results of APSD prove the effectiveness and feasibility of DALN.At the end of the paper,the research work of this article is summarized,and the future research direction is prospected. |
