Research On Fabric Defect Detection Method Based On Segmentation Network

Textile industry occupies a considerable proportion in Chinese industry,and its products are widely used in clothing,architecture and even aerospace.In the textile industry,the surface quality of products is an important factor that affects its price and grade evaluation.Traditional inspection methods use manual to detect surface defects,which is not only slow,but also can't guarantee the consistency of inspection results.In recent years,with the successful application of deep learning in various fields of images,it provides a new method for efficient and accurate detection of fabric defects.Semantic segmentation network can capture rich contextual information,and it can segment small target images relatively accurately and efficiently.It is suitable for small data sets,can use fewer samples for training and get good results,and is more suitable for fabric defect detection.Therefore,this paper uses semantic segmentation network to study the fabric defect detection technology.The main research contents are as follows:(1)A deep learning model based on two-level segmentation network is proposed.Firstly,a 9-layer full convolution network is realized by Re LU function and batch standardization technology to segment defects.Then,aiming at the problem that continuous downsampling will reduce the resolution of feature mapping and cause incomplete detail information,an empty space convolution pool pyramid module is added in the segmentation part.Finally,based on the characteristics of segmentation network,a decision network is proposed,and the segmentation results are used as input features to construct a decision network to classify defects.The method is tested on simple texture data set and complex texture data set.The experimental results show that the accuracy and recall rate of the method proposed in this paper exceed 90.5% and 95.4% for simple texture data set,and 89.2% and 86.1% for complex texture data set.(2)An improved fabric defect detection method based on U-Net segmentation network is proposed.Firstly,in view of the problem that the background interference of fabric defect image is too large,resulting in unsatisfactory segmentation effect,the U-Net network is taken as the network skeleton,and the attention module of feature pyramid is combined on the basis of it to better pay attention to the defect information,thus suppressing unimportant background information,and the structure and embedded position of attention module are redesigned.Secondly,in the aspect of loss function,because the model trained by a single cross entropy loss function will make the prediction result have a great background tendency,the cross entropy loss function and Dice coefficient loss function(dice)are combined for training.Finally,the experimental comparison of the proposed method on the fabric defect image data set is made.Experimental results show that the accuracy rate and recall rate of the proposed method for the adopted data set reach 77.8% and 85.4%,respectively,which are improved compared with the original network,thus proving the feasibility of the proposed method.At the same time,the segmentation results are visualized,and this method also achieves a good visualization effect on the defect target in the fabric image.