Research On Fabric Defect Detection Method Based On Improved Cascade R-CNN

Fabric defect detection is a crucial step in the fabric production process as defects can significantly impact the quality of the fabrics.However,many enterprises still rely on traditional manual inspection methods or methods based on traditional image processing for inspection.Manual detection is subject to the inspectors’ subjective factors,leading to fluctuating results,high costs,and low detection efficiency.On the other hand,the method based on traditional image processing relies heavily on feature design,which is affected by the designer’s experience and may not perform well in generalizing to new defects.In recent years,with the development of deep learning,it has achieved great success in the image field because of its ability to automatically extract features,and it has also become the main development trend of fabric defect detection.Cascade R-CNN is one of the state-of-the-art object detection networks.This thesis improves the fabric features,combines the frequency domain processing method,introduces the frequency domain features from both the macroscopic data processing and the microscopic network structure,and proposes a fabric defect detection method based on an improved Cascade R-CNN.This thesis mainly focuses on the following aspects:(1)Researching on data augmentation: Online and offline augmentation methods are adopted to solve the problem of insufficient sample size and uneven class distribution in the disclosed fabric defect detection dataset.This not only expands the dataset to enhance the model’s generalization ability,but also amplifies the frequency domain features of defects,so that the model is more able to learn the difference between the defect texture and the normal background,thereby improving the model’s detection accuracy.(2)Researching on deep learning network structures: When the model is downsampled to extract features,an additional frequency domain channel attention mechanism is introduced to expand the receptive field,so that the model can focus more on features that are more closely related to the object without losing feature diversity,thereby obtaining richer and more effective deep semantic information.(3)Researching on model fitness: To better adapt to the fabric defect dataset,the network has been modified based on its features.Appropriate multi-scale training is chosen based on the pixel characteristics of the dataset.The ratio and scale of the Anchors are adjusted according to the size of the fabric defects.The Soft-NMS algorithm is used instead of the traditional NMS algorithm for bounding box filtering.Ro I Align is used instead of Ro I pooling for a more precise feature extraction,making the improved network more effective for fabric defect detection.The experimental results show that the detection accuracy of the method proposed in this paper reaches 0.558 m AP in the public TIANCHI-XUELANGAI dataset,which is 9.2% higher than baseline.This method not only has a good detection effect,but also works in different basic models,and is a generally improved method for fabric defect detection.