Low-level image processing for salient object detection

Abstract

Human vision is an extraordinary system made up of complex structures and processes; one such mechanism is selective processing. Despite the mass of stimuli that meets the eyes every second (estimated between 108 and 109 bits), many complex tasks, such as object detection and facial recognition, are performed with ease. Imitating some of the characteristics of the human visual system has become a key focus for a range of research communities. The key characteristic of focus within this thesis is saliency detection, which aims to detect and segment areas/objects of interest while suppressing all other non-interesting information.

The first contribution of this thesis investigates the role of individual features when determining the salient regions of an image, with the aim of amalgamating the dominant features to produce a saliency model. Further to this, the proposed model is tasked with a linear classification problem, evaluating saliency features against traditional methods. One feature that is scarcely adopted amongst low-level saliency approaches is gradient magnitude. The impact of gradient magnitude as a saliency feature is the focus for the second contribution. A family of scalable derivative operators are implemented for use within the proposed saliency framework. The final contribution presents, a novel method to identify multiple scales within a single scale image, to aid in determining which regions are relevant.

The methodologies presented in this thesis result in a novel framework to detect salient objects within images. The evaluation of individual features provides a firm basis to design a single scale saliency algorithm, denoted as the Gradient Colour Contrast Saliency (GCCS) algorithm. Following this, saliency features are applied to a linear classification task, achieving a greater classification accuracy than the traditional HoG-SVM approach. With gradient adopted as a feature, the model is assessed when calculating the features using a family of scalable Gaussian based derivative operators. The introduction of the Near-Circular operator was found to further improve the model’s performance. Finally, a novel approach is developed to determine multiple processing scales within a single scale image. This method allows for image regions that have high activity (likely containing salient pixels) to be detected. From this basis, feature adaptivity is introduced to the framework. The size of the gradient operator applied at each image sub-region is dependent on the scale value of said region. Global colour only processes regions with a higher scale value than an algorithmically dependent threshold. This novel approach aligns closely with biological vision by only processing relevant information. The approach results in computational performance improvement and it outperforms a number of state-of-the-art saliency methods.

Date of Award	Nov 2022
Original language	English
Supervisor	Bryan Scotney (Supervisor), Bryan Gardiner (Supervisor) & Sonya Coleman (Supervisor)