Evolutionary Computation for Feature Manipulation in Salient Object Detection

Abstract

The human visual system can efficiently cope with complex natural scenes containing various objects at different scales using the visual attention mechanism. Salient object detection (SOD) aims to simulate the capability of the human visual system in prioritizing objects for high-level processing. SOD is a process of identifying and localizing the most attention grabbing object(s) of a scene and separating the whole extent of the object(s) from the scene. In SOD, significant research has been dedicated to design and introduce new features to the domain. The existing saliency feature space suffers from some difficulties such as having high dimensionality, features are not equally important, some features are irrelevant, and the original features are not informative enough. These difficulties can lead to various performance limitations. Feature manipulation is the process which improves the input feature space to enhance the learning quality and performance.

Evolutionary computation (EC) techniques have been employed in a wide range of tasks due to their powerful search abilities. Genetic programming (GP) and particle swarm optimization (PSO) are well-known EC techniques which have been used for feature manipulation.

The overall goal of this thesis is to develop feature manipulation methods including feature weighting, feature selection, and feature construction using EC techniques to improve the input feature set for SOD.

This thesis proposes a feature weighting method utilizing PSO to explore the relative contribution of each saliency feature in the feature combination process. Saliency features are referred to the features which are extracted from different levels (e.g., pixel, segmentation) of an image to compute the saliency values over the entire image. The experimental results show that different datasets favour different weights for the employed features. The results also reveal that by considering the importance of each feature in the combination process, the proposed method has achieved better performance than that of the competitive methods.

This thesis proposes a new bottom-up SOD method to detect salient objects by constructing two new informative saliency features and designing a new feature combination framework. The proposed method aims at developing features which target to identify different regions of the image. The proposed method makes a good balance between computational time and performance.

This thesis proposes a GP-based method to automatically construct foreground and background saliency features. The automatically constructed features do not require domain-knowledge and they are more informative compared to the manually constructed features. The results show that GP is robust towards the changes in the input feature set (e.g., adding more features to the input feature set) and improves the performance by introducing more informative features to the SOD domain.

This thesis proposes a GP-based SOD method which automatically produces saliency maps (a 2-D map containing saliency values) for different types of images. This GP-based SOD method applies feature selection and feature combination during the learning process for SOD. GP with built-in feature selection process which selects informative features from the original set and combines the selected features to produce the final saliency map. The results show that GP can potentially explore a large search space and find a good way to combine different input features.

This thesis introduces GP for the first time to construct high-level saliency features from the low-level features for SOD, which aims to improve the performance of SOD, particularly on challenging and complex SOD tasks. The proposed method constructs fewer features that achieve better saliency performance than the original full feature set.