Research On Fabric Defect Method Based On Deep Learning

Fabric defect detection is an essential in the textile industry.During the production of fabrics,different kinds of defects are inevitably produced.Therefore,accurate and fast fabric defect detection is of great significance to improve the production efficiency of textile enterprises.At present,most textile enterprises mainly rely on artificial vision for defect detection,which is inefficient and labor-intensive.The fabric defect detection method using deep learning has been significantly improved compared to the traditional defect detection method,but fabric defects still have problems such as complex color background interference,small objects,extreme aspect ratios,and unbalanced classification,which will lead to the frequent occurrence of low defect detection recognition rate,false detection,and missed detection.In view of the above problems,this paper proposes research on the detection method of fabric defects based on deep learning.The main research contents are as follows:A fusion of deformation convolution and self-attention for plain fabric defect detection is proposed.First,multi-scale feature extraction with fused deformation convolution is proposed,which can effectively alleviate the problem of the model's insufficient ability to extract irregular defect features.Second,a dual-channel feature fusion is proposed,and more effective defect features are obtained by introducing self-attention mechanism adaptive adjustment and fusion,so as to generate a new defect feature map with strong semantics and precise location information,which improves the accuracy of defect detection of many small objects in the fabric defect data set.Finally,an adaptive bounding box generator is designed in the region proposal network to guide the design of the initial anchor box to obtain a more accurate object bounding box,which solves the problem that some defects in the fabric defect dataset(such as hanging warps)with different aspect ratios cannot generate tight bounding boxes.The experimental comparison shows that the proposed method has a good detection effect,and effectively improves the accuracy and efficiency of the defect detection of plain fabrics.A two-branch high-resolution transformer network for fabric defect detection method is proposed.A two-branch high-resolution feature extraction is proposed in the feature extraction stage to eliminate the influence of complex floral backgrounds and extract clearer and more representative defective features.Reduces the effect of fancy backgrounds in fabric imperfections on feature extraction and removes background noise unrelated to the target imperfections.Designing multi-scale feature pyramid transformer to extract multiscale features across space and scale.Then design an adaptive bounding box generator in the region proposal network to facilitate subsequent detection and regression.Finally,the use of an improved focal loss reduces the sensitivity of the model to the lack of defective samples and solves the problem of poor detection accuracy for a few classes of defective samples.The experimental results show that this method can effectively improve the detection accuracy of fabric defects,which is better than the current mainstream fabric defect detection methods.