Abstract:In this paper, we propose a new approach to perform supervised texture classification/segmentation. The proposed idea is to feed a Fully Convolutional Network with specific texture descriptors. These texture features are extracted from images by using an empirical curvelet transform. We propose a method to build a unique empirical curvelet filter bank adapted to a given dictionary of textures. We then show that the output of these filters can be used to build efficient texture descriptors utilized to finally feed deep learning networks. Our approach is finally evaluated on several datasets and compare the results to various state-of-the-art algorithms and show that the proposed method dramatically outperform all existing ones.

What problem does this paper attempt to address?

### What problem does this paper attempt to solve? This paper aims to solve the challenging problems in supervised texture classification/segmentation. Specifically, the author proposes a new method to improve the classification and segmentation accuracy of texture images by combining the Empirical Curvelet Transform (EWT2DC) and the Fully Convolutional Network (FCN). #### Research background and motivation Texture analysis is an important aspect in computer vision and is widely used in fields such as medical imaging (such as tumor detection), camouflage target detection, and microscopic imaging (such as molecular arrangement analysis). However, due to the high variability of textures and the difficulty in accurately describing them with mathematical models, texture segmentation and classification remain an open and challenging problem in the fields of image processing and computer vision. Existing texture classification/segmentation algorithms are usually divided into two main steps: 1. **Feature extraction**: Extract features that can represent textures from the image. 2. **Classification**: Use this feature information to distinguish different textures. In the case of supervised learning, the algorithm is trained using a known texture dictionary to learn how to distinguish different textures. This paper focuses on supervised classification/segmentation and proposes a new method that uses data - driven wavelets (i.e., empirical curvelets) to create efficient texture descriptors, which are then used in a deep - learning network for classification. #### Proposed method To overcome the shortcomings of existing methods in extracting low - level texture features, especially their weaknesses in the face of different lighting conditions or similar geometric structures, the author proposes to combine the FCN architecture with empirical wavelet features. The specific steps are as follows: 1. **Empirical Curvelet Transform (EWT2DC)**: Construct a set of unique empirical curvelet filters from a given texture dictionary. 2. **Feature extraction**: Apply these filters to extract texture features. 3. **FCN classification**: Input the extracted features into the FCN for pixel - level classification. Through this method, the author hopes to significantly improve the segmentation and classification effects on multiple texture datasets. Experimental results show that the proposed fully convolutional network based on the empirical curvelet transform outperforms various existing advanced algorithms on multiple standard datasets. ### Formula summary - **Empirical Curvelet Transform (EWT2DC)**: - Definition of low - pass filter: \[ F_2(\phi_1)(\omega, \theta)= \begin{cases} 1 & \text{if }|\omega|\leq(1 - \gamma)\omega_1,\\ \cosh\left(\frac{\pi}{2\beta}\left(\frac{1}{2\gamma\omega_1}(|\omega|-(1 - \gamma)\omega_1)\right)\right) & \text{if }(1 - \gamma)\omega_1\leq|\omega|\leq(1 + \gamma)\omega_1,\\ 0 & \text{otherwise} \end{cases} \] - Definition of band - pass filter: \[ W_n(\omega)= \begin{cases} 1 & \text{if }(1 + \gamma)\omega_n\leq|\omega|\leq(1 - \gamma)\omega_{n + 1},\\ \cosh\left(\frac{\pi}{2\beta}\left(\frac{1}{2\gamma\omega_{n + 1}}(|\omega|-(1 - \gamma)\omega_{n + 1})\right)\right) & \text{if }(1 - \gamma)\omega_{n + 1}\leq|\omega| \end{cases} \]

Empirical curvelet based Fully Convolutional Network for supervised texture image segmentation

Multi-Level Feature Descriptor for Robust Texture Classification via Locality-Constrained Collaborative Strategy

Rough-fuzzy set approach for color and texture based image segmentation

Texture segmentation with Fully Convolutional Networks

Unsupervised Texture Segmentation Based on Geodesic Active Regions

Face Recognition Based on Curvelets and Local Binary Pattern Features Via Using Local Property Preservation

Unsupervised Image Segmentation Based On Contourlet Texture Features And Byy Harmony Learning Of T-Mixtures

Binary patterns encoded convolutional neural networks for texture recognition and remote sensing scene classification

Learning Statistical Texture for Semantic Segmentation

weghts Feature Extractor Siamese Texture Encoder Decoder 3 × 3 Conv GCM Relation Network DirConv Unit DAM : Directionality-Aware Module GCM : Global Context Module Concat Operation Channel-Wise Attention

Fusion of Complex Networks-based Global and Local Features for Texture Classification

MedFCT: A Frequency Domain Joint CNN-Transformer Network for Semi-supervised Medical Image Segmentation

Unsupervised Texture Segmentation Via Applying Geodesic Active Regions To Gaborian Feature Space

Deep convolutional filter banks for texture recognition and segmentation

Texture Segmentation Using Joint Time Frequency Representation and Unsupervised Classifier

Weakly-Supervised Sparse Coding with Geometric Prior for Interactive Texture Segmentation

Fast Texture Segmentation Based on Semi-local Region Descriptor and Active Contour Driven by the Bhattacharyya Distance

Review of wavelet-based unsupervised texture segmentation, advantage of adaptive wavelets

Encoding Spatial Distribution of Convolutional Features for Texture Representation.

A Segmentation Method of Color Texture Image

CDF Space Covariance Matrix of Gabor Wavelet with Convolutional Neural Network for Texture Recognition.