Research On The Saliency Detection Method Of Textile Defects Fused With Information Entropy

Fabric defect detection is an important part of the textile production process.There are many disadvantages to manual detection,which makes it difficult to meet production requirements.With the development of machine vision and image processing technology,automatic detection of fabric defects began to replace manual detection.However,many fabric defect detection methods are less adaptable to different texture patterns and defect types.Therefore,this study uses information entropy to represent texture features.Imitate the characteristics of human visual attention mechanism that can quickly and accurately locate the region of interest to complete the location and extraction of fabric defects.First,we use information entropy to describe the texture of the fabric image.The texture of the defect-free region is deterministic and periodic.When the fabric is defective,the texture of the defective region is random.The one-dimensional entropy of the image is a measure of the uncertainty and randomness of the texture.Based on the one-dimensional entropy,two-dimensional entropy that reflects the spatial distribution of the texture is introduced to distinguish the texture difference between the defect-free region and the defect region.Then,imitating the human visual system's attention mechanism to salient features.We use quaternions to re-represent the fabric image.The representation method of RGB color space is not suitable for the human visual mechanism.In the classic visual attention model,the opposite color space is used to represent the image.Based on the characteristics of fabric images,this paper combines the two-dimensional entropy describing textures with salient features in opposite color spaces to form four channels,and uses the ability of quaternions to represent multi-channel images.An improved quaternion image representation is proposed.The method fully extracted the texture information and saliency features of the fabric image.Finally,the quaternion image is transformed into the frequency domain by using a Hypercomplex Fourier transform.By analyzing the characteristics of the fabric image frequency spectrum,its amplitude spectrum is locally tuned to preserve the salient information including defects and suppress non-salient information that includes background patterns.The saliency map of the defect region is obtained by inverse transforming the tuned spectrum information.For saliency maps,Otsu threshold method can be used to segment defects and mark them.Experiments are performed using a public database of fabric images.The experimental results show that the method has a good detection effect on different defects in various fabric patterns.The defect detection rate can reach 98%.In addition,the image-level and pixel-level indicators were used to evaluate the research method and several excellent fabric defect detection methods.The results show that the research method has advantages in most indicators and comprehensive evaluation,which proves the research method can meet the requirements of on-line detection of fabric defects.It has a good guiding significance for defect detection on other surfaces.