Multi Resolution and level Set Based Approach For texture Segmentation

Abstract

Texture segmentation of images has emerged as a prominent field of research in computer vision The primary goal of segmentation is to classify comparable groupings of pixels based on unique properties, thus minimizing processed data. This approach is useful in a variety of applications, including urban traffic management, surveillance, medical diagnosis, military target detection, environmental monitoring, and item tracking. Managing the rich textures seen in normal and crosshatched photographs becomes a significant problem. These textures include brightness, homogeneity, density, roughness, consistency, linearity, frequency, phase, orientation, coarseness, softness, and granulation as local sub-pattern qualities. Given that texture is a visually complex substance, developing effective features capable of completely analyzing various textures and detecting qualities separating them is a significant challenge. While previous decades have witnessed numerous research works by employing methods like filter banks, co-occurrence statistics, and Hidden Markov Models (HMMs) but still there are lot of limitations for precise texture segmentation. To overcome these challenges and facilitate optical and crosshatched texture segmentation, a novel technique termed the Multi-Resolution Feature Embedded Level Set Model (MRFE-LSM) has been proposed. newlineIn the first study, to overcome the above mentioned difficulties, a novel approach called MRFE-LSM is introduced for texture segmentation in normal images. This method employs frequency domain filter-based texture feature extraction with precise boundary identification using level sets. The initial step involves applying a [2×2] low-pass Gaussian filter to the original image, reducing high-frequency transitions and intensity variations. The efficacy of MRFE-LSM is validated using the Brodatz database and a customized Microsoft Office dataset. Results indicate that the proposed method effectively segments the Regions Of Interest (ROIs), achieving Intersection Over Union (IOU) parameter coefficients surpassing 0.8 and maintaining strong alignment with the original texture image. newlineIn the second study, a novel MRFE-LSM-Modified Support Vector Machine (MSVM) approach is proposed for crosshatched texture segmentation. By using the Brodatz texture dataset, this method enhances the distinction between low-intensity variation regions in crosshatched textures using multi-resolution features from frequency domain filters. The subsequent step involves addressing crosshatched texture newlinev newlinesegmentation using level set-based active contour models. Experimental evaluations on this method reveal its success in precisely segmenting the ROIs in line with the original image. Comparing the performance with the benchmark Modified Convolutional Neural Network with Whale Optimization Algorithm (MCNN-WOA) model, the suggested MRFE-LSM-MSVM method exhibits superior accuracy in crosshatched texture segmentation. newlineIn the third study, a deep analysis about MRFE-LSM-MSVM method is presented for crosshatched texture segmentation. A comprehensive investigation is conducted employing 5-fold and 10-fold cross-validation, with and without histogram inclusion, to demonstrate the segmentation performance. Additionally, a comparison is drawn between the MRFE-LSM-MSVM model and an existing feature segmentation model, MCNN- WOA. The MCNN-WOA achieves impressive metrics on the Brodatz dataset, including 99.71% accuracy, 96.70% precision, 95.80% recall, and a 96.20% F1-score. However, the MRFE-LSM-MSVM model outperforms with an F1-score of 98.90%, 99.82% accuracy, 99.12% precision, and 98.88% recall. Notably, the suggested model excels in accurately segmenting textures and also addresses the challenge of long processing time, highlighting its distinct advantage. newline