A SYSTEM FOR RETRIEVAL OF TEXTURE VISION USING ADAPTIVE TETROLET TRANSFORMS

Abstract

The present invention relates to a system for retrieval of texture vision using adaptive tetrolet transforms. The system comprises: a data collection module comprises 116 images sourced from the Brodatz database, 40 images from the Vision Texture database, 91 texture images which are rotated at different angles and collected from the texture dataset; feature extraction module utilizing standard deviation and mean to create feature vectors, and employed Tetrolet transform. The present system excels in capturing fine texture details through a unique analysis approach. By applying tetrominoes at each decomposition level, we select the optimal combinations that best represent the geometric structure of the images. The high-pass components obtained from the tetrolet decomposition serve as the basis for feature computation, where a feature-set is constructed by combining the standard deviation and energy of high pass filter at different levels of decomposition. The performance of the proposed invention is evaluated on benchmark image databases, specifically Brodatz and VisTex. The test results demonstrate that our approach attains retrieval accuracies of 78.80% for group D1 (Brodatz), 84.41% for group D2 (VisTex), and 77.41% for group D3 (rotated images from Brodatz), highlighting the effectiveness of tetrolet transforms in enhancing texture image retrieval.

Specification

Description:Field of the Invention
The invention relates to an image texture retrieval method, more particular to a system for retrieval of texture vision using adaptive tetrolet transforms, which excel in capturing fine texture details through a unique analysis approach.
Background of the Invention
The invention outlined in this context focuses on addressing the challenges of efficient and effective image retrieval in large digital image databases, especially when handling texture-rich images. With the exponential growth of multimedia content, particularly in digital image libraries, traditional Text-Based Image Retrieval (TBIR) methods have proven ineffective due to several limitations. Here, the focus is on improving the Content-Based Image Retrieval (CBIR) techniques, particularly for texture image retrieval, which is essential in fields like medical imaging, remote sensing, and material science. Roland Kwitt et al. [2011] employed Copulas within a Bayesian framework for texture image extraction, achieving competitive retrieval accuracy and efficient runtime performance, making it suitable for large-scale image databases. Their study also includes a detailed computational analysis covering runtime measurements, storage requirements, and the core components of the approach. In [Nour-Eddine Lasmar et al.(2014)], a Gaussian Copula-based multivariate modeling approach was introduced for texture image retrieval, which decouples the dependence structure from the marginal distributions. The authors proposed two new models using generalized Gaussian and Weibull densities, respectively, and integrated Gaussian Copula modeling with Jeffrey divergence to define a similarity measure. In [Ming Hong Pi et al. (2006)], a novel method is proposed for composing image features based on the decomposed subbands. The approach utilizes "three-pass layer probability (TPLP)" and bit plane signatures, offering the advantage of low computational complexity, as it eliminates the need for dequantization and feature extraction from wavelet coefficients.
Challenges and solution in CBIR Systems:
Higher Accuracy: A key challenge in CBIR is achieving high retrieval accuracy, especially in large, heterogeneous datasets. This challenge is addressed by achieving the higher accuracy in retrieving the texture images from the heterogeneous and larger datasets of texture images
Feature Extraction: The quality of the feature extraction process (e.g., using color, texture, and shape) directly impacts the system's ability to retrieve relevant images. The extracted features are of higher quality due to the adaptive tetrolet transform which extracts the texture feature by analysing the texture images in 117 different positions of tetromenoes.
Computational Efficiency: The need to balance high precision with low computational cost is crucial for a CBIR system to be scalable and usable in real-time applications. The proposed invention achieves the less computation cost by spending the less time and space complexity in the storage and retrieval of texture features.
Drawings
Figure 1: The Proposed Invention
Figure 2. Performance on database D1
Figure 3. Performance on database D2
Detailed Description of the Invention
The following description includes the preferred best mode of one embodiment of the present invention. It will be clear from this description of the invention that the invention is not limited to these illustrated embodiments but that the invention also includes a variety of modifications and embodiments thereto. Therefore, the present description should be seen as illustrative and not limiting. While the invention is susceptible to various modifications and alternative constructions, it should be understood, that there is no intention to limit the invention to the specific form disclosed, but, on the contrary, the invention is to cover all modifications, alternative constructions, and equivalents falling within the spirit and scope of the invention as defined in the claims.
In any embodiment described herein, the open-ended terms "comprising," "comprises," and the like (which are synonymous with "including," "having" and "characterized by") may be replaced by the respective partially closed phrases "consisting essentially of," consists essentially of," and the like or the respective closed phrases "consisting of," "consists of, the like. As used herein, the singular forms "a", "an", and "the" designate both the singular and the plural, unless expressly stated to designate the singular only.
The present invention invention aims to improve retrieval accuracy while minimizing feature extraction time in Content-Based Image Retrieval (CBIR) systems. The performance of these systems is primarily determined by two factors: retrieval accuracy and the time required to find similar images. The quality of the feature vector is crucial for achieving higher retrieval accuracy, while the size of the feature vector and the efficiency of the feature extraction process contribute to reduced retrieval time.
Each image in the database undergoes decomposition using Tetrolet transforms, analyzed up to the fourth level of decomposition. A feature vector is generated by calculating the standard deviation and energy at each level separately. These two values are then combined to form the feature vector for each image in the database.
The choice to use energy as a texture feature stems from its ability to represent energy distribution in the frequency domain, which effectively characterizes the texture of the image. This approach allows for a more detailed and meaningful representation of image content, enhancing retrieval accuracy in the CBIR system.
Algorithm for Image Retrieval
Input:Image
Output: Group of Similar images
1. Loading and Preprocessing of the query image:
o Load the image of size 512×512.
o If the image is in RGB format, convert it to grayscale.
2. Divide the Image:
o To split an image into 16 subimages, we divide the image into a 4x4 grid, which means the image is cut into 4 rows and 4 columns.
3. Adaptive feature Extraction
o Perform Tetrolet transform on the image using all 117 possible combinations, up to the 4th level of decomposition.
o Arrange the high and low pass subbands according to equations (5) and (6).
o Use a bijective mapping method for subband representation.
4. Select Best Tile:
o At each decomposition level, select the best tile from the 117 combinations, ensuring optimal local geometry representation.
5. Calculate Statistical Features:
o For each decomposition level, compute the standard deviation along with the mean for the 3 frequency subbands of high pass filter.
6. Construct Feature Vector:
o Form the feature set by adding the mean and standard deviation value for each high pass filter.
7. Compute Euclidean Distances:
o Determine the Euclidean distance between the feature-set of the requested reference image and the feature vectors of all stored database images.
8. Rank Distances:
o Sort the distances in increasing order to identify the most similar images.
9. Retrieve Relevant Images:
o Sort the images in the series of their relevance to the requested image, based on the distance from the query image.
By following this algorithm, the system efficiently retrieves images that are most relevant to the input query image based on the extracted feature vectors as shown in Fig 1.
Experimental Setup and Result Analysis:
To evaluate the image retrieval performance of the proposed invention, tests were conducted using three distinct benchmark databases. Below are the details of each database:
Database D1:
o This texture database comprises 116 images sourced from the Brodatz database. Each image is divided into non-touching regions of size 128×128, resulting in 1,856 images.
Database D2:
o The texture database D2, as used in [20], contains 40 images from the Vision Texture database. Similar to Database D1, each image is partitioned into the non-intersecting block/region of the size of 128×128, leading 640 sub-images.
Database D3:
o This database comprises of 91 texture images which are rotated at different angles and collected from the texture dataset []. Images in this database are rotated at angles of 0°, 30°, 60°, 90°, 120°, 150°, and 180°. Each image is sized 512×512 and is divided into sixteen non-touching regions of 128×128, creating 1,456 images.
Experiments Conducted
Two primary experiments were performed using Databases D1 and D2:
• First Set of Experiments:
o This experiment utilized Database D2, where Manjunath and Ma employed Gabor wavelet transform for texture image retrieval, utilizing standard deviation and mean to create feature vectors.
• Second Set of Experiments:
o Again using Database D2, we compared our results with those of Do and Vetterli , who implemented spectral domain based texture image retrieval by merging generalized Gaussian density with the Kullback-Leibler distance method.
Additionally, we compared our findings with the work of M. Kokare et al. [6], who introduced the extraction of rotated textures images as well using rotated complex wavelet filters that enhances the retrieval system in terms of both optimized time and higher accuracy. They also used Databases D1 and D2 with similar image counts.
Performance on Rotated Images
Given the properties of the Tetrolet transform, which accounts for reflections and all possible rotations, we conducted a new experiment on the rotated texture image database (Database D3). Our proposed method demonstrated strong performance, achieving a high retrieval accuracy of 78.41% for these rotated images. This validates the effectiveness of our approach in handling variations in image orientation.



This invention is going to offer the following advantages:
? Wavelet Transform and Multiresolution Analysis: This technique provides multiresolution analysis of an image, which helps capture both color and texture features. The use of wavelet transforms allows for the decomposition of the image into various frequency components, providing detailed insights at different resolutions.
? Adaptive texture features: The paper introduces Tetrolet Transform, a specialized form of the Haar Wavelet, to analyze the texture of images. This transform is particularly well-suited for extracting texture features from images, making it highly effective in texture-based CBIR systems due to its higher adaptive and dynamic feature extraction process.
? Rotation Invariance: One of the key challenges in texture image retrieval is dealing with image rotations and reflections. Tetrolet transform addresses this issue by analyzing the image under all possible rotations and reflections (a total of 117 possibilities). By selecting the best tile from these options, the system becomes rotation-invariant, ensuring that texture features are consistently captured regardless of the image geometry.
? The Proposed Feature vector can perform well on the texture images.
? This invention improves the overall performance of the CBIR system by reducing the search time, feature extraction time along with the reduced space to store the feature vector.
, Claims:1. A system for retrieval of texture vision using adaptive tetrolet transforms, comprises of:
a data collection module comprises 116 images sourced from the Brodatz database, 40 images from the Vision Texture database, 91 texture images which are rotated at different angles and collected from the texture dataset; feature extraction module utilizing standard deviation and mean to create feature vectors, and employed Tetrolet transform.
2. The system for retrieval of texture vision using adaptive tetrolet transforms as claimed in the claim 1, wherein process for image retrieval comprises the following steps:
o Step 1: Load the image of size 512×512, If the image is in RGB format, convert it to grayscale;
o Step 2: split an image into 16 subimages, divide the image into a 4x4 grid, which means the image is cut into 4 rows and 4 columns.
o Step 3: Perform Tetrolet transform on the image using all 117 possible combinations, up to the 4th level of decomposition. Arrange the high and low pass subbands and use a bijective mapping method for subband representation.
o Step 4: At each decomposition level, select the best tile from the 117 combinations, ensuring optimal local geometry representation;
o Step 5: For each decomposition level, compute the standard deviation along with the mean for the 3 frequency subbands of high pass filter;
o Step 6: Form the feature set by adding the mean and standard deviation value for each high pass filter;
o Step 7: Determine the Euclidean distance between the feature-set of the requested reference image and the feature vectors of all stored database images;
o Step 8: Sort the distances in increasing order to identify the most similar images; and
o Step 9: Sort the images in the series of their relevance to the requested image, based on the distance from the query image.
3. The system for retrieval of texture vision using adaptive tetrolet transforms as claimed in the claim 1, wherein evaluate the image retrieval performance of the proposed invention, tests were conducted using three distinct benchmark databases:
o D1 Dataset: texture database comprises 116 images sourced from the Brodatz database. Each image is divided into non-touching regions of size 128×128, resulting in 1,856 images.
o D2 Dataset: The texture database D2, contains 40 images from the Vision Texture database. Similar to Database D1, each image is partitioned into the non-intersecting block/region of the size of 128×128, leading 640 sub-images; and
o D3 Dataset: This database comprises of 91 texture images which are rotated at different angles and collected from the texture dataset. Images in this database are rotated at angles of 0°, 30°, 60°, 90°, 120°, 150°, and 180°. Each image is sized 512×512 and is divided into sixteen non-touching regions of 128×128, creating 1,456 images.
4. The system for retrieval of texture vision using adaptive tetrolet transforms as claimed in the claim 1, wherein method demonstrated strong performance, achieving a high retrieval accuracy of 78.41% for these rotated images.

Documents