Multi-Resolution Feature Embedded Level Set Model for Crosshatched Texture Segmentation

Abstract

In image processing applications, texture is the most important element utilized by human visual systems for distinguishing dissimilar objects in a scene. In this research article, a variational model based on the level set is implemented for crosshatched texture segmentation. In this study, the proposed model’s performance is validated on the Brodatz texture dataset. The cross-hatched texture segmentation in the lower resolution texture images is difficult, due to the computational and memory requirements. The aforementioned issue has been resolved by implementing a variational model based on the level set that enables efficient segmentation in both low and high-resolution images with automatic selection of the filter size. In the proposed model, the multi-resolution feature obtained from the frequency domain filters enhances the dissimilarity between the regions of crosshatched textures that have low-intensity variations. Then, the resultant images are integrated with a level set-based active contour model that addresses the segmentation of crosshatched texture images. The noise added during the segmentation process is eliminated by morphological processing. The experiments conducted on the Brodatz texture dataset demonstrated the effectiveness of the proposed model, and the obtained results are validated in terms of Intersection over the Union (IoU) index, accuracy, precision, f1-score and recall. The extensive experimental investigation shows that the proposed model effectively segments the region of interest in close correspondence with the original image. The proposed segmentation model with a multi-support vector machine has achieved a classification accuracy of 99.82%, which is superior to the comparative model (modified convolutional neural network with whale optimization algorithm). The proposed model almost showed a 0.11% improvement in classification accuracy related to the existing model.