Research On Texture Classification Guided By Structural Semantics

In the field of computer vision,texture features play a crucial role in images as the basis of tasks such as image recognition,semantic segmentation and image synthesis.Affected by noise,content scale and other factors,the effective feature extraction of complex texture images is an important reason for affecting the classification accuracy.Artificial feature methods are usually efficient to extract texture structures based on subjective understanding,but may ignore non-intuitive textures.Convolutional neural networks can automatically search for texture features including non-intuitive textures according to the task,and can achieve good performance for specific classification tasks.However,the existing algorithms have shortcomings in maintaining texture details and capturing long-term correlation of textures,and the description of intuitive texture structures such as periodic repetition is not as effective as manual feature methods.Therefore,for texture classification tasks,this paper introduces an artificial feature module into the convolutional neural network architecture to take into account both intuitive and non-intuitive texture features.The details are as follows:(1)The low-level detail features and various scale features of texture images play different roles in image identification.Traditional deep neural networks focus more on the high-level semantic representation of texture images,and have certain limitations in preserving the local details of texture.In order to focus directly to the specified scale features more conveniently and effectively,the wavelet transform is considered.Wavelet transform can decompose an image into frequency subbands with different scales and directions,and texture details of different scales can be easily obtained by decomposing the image layer by layer.Therefore,this paper introduces Haar wavelet transform to preserve texture details in limited direction and limited scale,so as to achieve higher accuracy of texture classification.Aiming at the problem of scale selection for wavelet transform,this paper proposes an adaptive scale selection rule based on the maximum difference of energy by using the spatial energy distribution characteristics of texture images at different scales.The rule is applied to drive the network to select the decomposition depth according to the texture distribution,which avoids the excessive smoothing of lowfrequency components to weaken the comprehensive expression ability of the network while maintaining the detail features as much as possible.Experimental results on five texture classification datasets support the effectiveness of the proposed method in preserving texture details.(2)The global texture of many images is composed of approximately repeated local structures,which imply long-range dependencies.It is difficult for existing convolutional neural networks to effectively capture such long-range dependencies.In view of the fact that fractal dimension can easily measure self-similarity and long-range dependence and is insensitive to scale and rotation,in order to enhance the capture of longrange dependence,this paper introduces the statistical geometry fractal coding module into the convolutional neural network.This module quantifies the spatial arrangement pattern of texture by local and global fractal analysis,and realizes the description of the spatial distribution of texture.The performance of the classification network with fractal coding on the benchmark data set is improved to a certain extent compared with the benchmark network,which verifies the effectiveness of the proposed method and enhances the interpretability of the network.(3)Texture often presents obvious periodicity in spatial distribution,which can be more intuitive and easy to quantify in the frequency domain.Although the deep neural network method has the implicit ability to learn frequency domain features,it is inefficient and interpretable.In order to make more effective use of the frequency domain characteristics of texture and compensate for the shortcomings of convolutional neural network,a two-stream texture classification network with complementary spatial domain and frequency domain is designed.The effectiveness of the proposed network is demonstrated in comparison with a baseline classification network.The above three methods are combined to compare with the current advanced texture classification methods on DTD and FMD datasets,and the necessary ablation experiments are done.Experiments show that the proposed method effectively improves the performance of texture classification and shows good compatibility.