IMPROVED SALIENCY OF CLUTTERED IMAGES USING STRUCTURE EXTRACTION FROM TEXTURE

Abstract

Automatic salient region detection in an image is a useful technique that can assist in many computer vision tasks such as image segmentation and object recognition. It allows processing of input images without prior information of its contents. In this thesis work, a salient object detection approach which aims in limiting the distractions caused by small patterns in the background and foreground of an image is considered. First, a structure extraction algorithm is employed to smooth the textures present in the input image while preserving its structure information. Second, the structure image is segmented into perceptually homogenous regions by using graph-based image segmentation. Third, saliency map is computed according to color contrast and complimentary priors. The performance of the proposed method is compared against the prior method (without structure extraction) with the help of salient object detection datasets. Quantitative evaluation of the saliency maps are conducted using receiver operating characteristics (ROC) curve, overlap ratio (OR), weighted F- measure score and structure similarity (SSIM) index. The proposed method obtained a weighted-F score of 0.643 as compared to the 0.607 obtained by the prior method in the case of Microsoft Research Asia (MSRA10k) dataset. The same evaluation measure was improved from 0.509 to 0.534 in the case of Extended Complex Scene Saliency Dataset (ECSSD) dataset. Similarly, the SSIM index score obtained was 0.751 and 0.648 as compared to 0.539 and 0.516 obtained by the prior method for each datasets respectively. The comparison of the proposed method with three other existing salient region detection methods is also shown. The proposed method obtained a competitive score for MSRA10k dataset images whereas it obtains the highest score considering OR, weighted-F measure and SSIM index in the case of ECSSD dataset with a score of 0.551, 0.534 and 0.648 respectively.