Research On Fabric Defect Detection Algorithm Via Deep Learning Based Visual Saliency Model

Fabric defect detection is the key link of textile quality monitoring.The automatic fabric detection method based on image processing and machine vision has become the first choice of textile enterprises due to its advantages such as fastness and high accuracy.As the defects and textures in the fabric have complexity and diversity,the traditional detection method has a low detection rate and poor adaptation.Human visual mechanism can quickly locate salient objects,so fabric defect detection based on visual saliency has great research value.Based on the powerful feature extraction capabilities of deep learning,this thesis innovates in deep learning models,attention mechanisms,etc.,and researches fabric defect detection algorithms via deep learning based visual saliency model.The research results are summarized as follows:1)A visual saliency model based on full convolutional network and convolutional attention mechanism for fabric defect detection was proposed.Aiming at the problem of low detection accuracy of existing fabric defect detection methods,the full convolutional network was adopted to extract multi-level and multi-scale features,improving the ability to characterize fabric texture.Then,an attention mechanism module is added to the side-output of backbone network to assign weights to different feature maps and different pixels of the same feature map,which further improves the effectiveness of feature extraction.Finally,the multi-level saliency maps were fused by a series of short connection structures to detect the defect accurately.The experimental results show that the proposed algorithm is superior in locating and detecting irregular defects than other salient detection methods.2)A visual saliency model based on multi-scale context-aware for fabric defect detection was proposed.FCN has the problems of information loss and small target cannot be reconstructed in the process of extracting features.To solve this problem,a multi-scale context-aware module composed of dilated convolution and ordinary convolution was used to simulate the receptive field of the human visual system,and strengthen the deep features learned by the model.At the same time,compared with using a single size convolution kernel structure,the model can obtain different scale features by utilizing multiple convolution kernels of different sizes.By fusing multi-level features,the final detection result is more accurate than the same convolution kernel.The experimental results show that the proposed algorithm has high accuracy,whether it is a large target or a small.3)A visual saliency model based on deep separable convolution and self-attention mechanism for fabric defect detection was proposed.The location of fabric defects is random,so the way to learn each pixel or channel of the image will bring huge computational pressure.Therefore,this method reduces the complexity of the model by replacing the convolution which has a large amount of computation with the deep separable convolutional layer.Then,the self-attention model is applied to obtain the global dependency of the image by learning the relationship between pixels in the image,and avoid the calculation of the channel and spatial features.The experimental results show that the model parameters are greatly reduced on the premise of maintaining detection for various types of defects.The research results of this paper improve the adaptability and detection effect of the existing fabric defect detection methods.The proposed method can locate and detect the defects of plain weave fabrics with high detection accuracy.The related algorithms can also be applied to the detection of surface defects such as glass,steel,and films,which have extremely high value of research.