| Surface defects,as one of the most common product problems in industrial manufacturing,not only affect the appearance of the product,but also reduce its reliability and service life.Therefore,an efficient defect detection solution is of great practical importance to improve product quality.Industrial defect detection based on deep learning has become one of the research hotspots in industry and academic fields for its non-contact nondestructive detection,low cost and high intelligence.Data-driven deep learning relies on large valid datasets.Due to the influence of industrial production,defect datasets often have problems such as insufficient training samples and imbalanced categories.In addition,the deployment cost of deep learning platform and other objective factors cannot achieve the popularization and application of industrial embedded defect detection technology.In view of the above problems,our paper proposed an effective lightweight small sample defect detection method,the specific contents are as follows:(1)To solve the problem of low detection accuracy caused by imbalanced small sample defect data,GT-CutMix offline data augmentation algorithm based on precise sampling of defect location and Random-CutMix augmentation algorithm under the condition that data only contains category label were proposed.Experiments were conducted in two aspects,classification task and object detection task,to verify the effectiveness of the offline data enhancement algorithm.The results show that both algorithms increase the diversity of samples and effectively solve the sample imbalance problem.Meanwhile,GT-CutMix data augmentation algorithm can fully preserve the texture and geometric features of defects.(2)In order to improve the generalization performance of the model,our paper optimized and improved the network structure from two aspects: the benchmark network for defect detection and the design of DSSD model,respectively.First,the channel shuffle unit was introduced on the basis of MobileNetv2 to increase the mobility and interaction ability of feature information within the network,which constitutes a lightweight S-DSSD defect detection model.Second,the PSA feature pyramid attention mechanism was introduced into S-DSSD to construct the SA-DSSD model,which realizes channel domain feature rescaling,enhances the sensitivity of the model to defect features,suppresses the influence of invalid background information,and thus further improves the defect detection performance.In our paper,the algorithms are validated on the X-SDD small sample defect dataset,and the detection accuracy of the proposed S-DSSD model is 76.83% and the detection rate is 45 FPS under the GT-CutMix data augmentation algorithm.the SA-DSSD model achieves 78.63% detection accuracy mAP and 40 FPS detection rate at the expense of some model complexity.Both proposed defect detection algorithms outperform existing models and provide selective solutions for focused requirements such as high speed rate and high accuracy in different industrial production scenarios.The research results in our paper solve the problem of imbalance between defect detection sample classes,improve detection accuracy while reducing model complexity,and promote the development of small sample defect detection technology. |
