Genyun Sun

Unsupervised image saliency detection with Gestalt-laws guided optimization and visual attention based refinement

Abstract Visual attention is a kind of fundamental cognitive capability that allows human beings to focus on the region of interests (ROIs) under complex natural environments. What kind of ROIs that we pay attention to mainly depends on two distinct types of attentional mechanisms. The bottom-up mechanism can guide our detection of the salient objects and regions by externally driven factors, i.e. color and location, whilst the top-down mechanism controls our biasing attention based on prior knowledge and cognitive strategies being provided by visual cortex. However, how to practically use and fuse both attentional mechanisms for salient object detection has not been sufficiently explored. To the end, we propose in this paper an integrated framework consisting of bottom-up and top-down attention mechanisms that enable attention to be computed at the level of salient objects and/or regions. Within our framework, the model of a bottom-up mechanism is guided by the gestalt-laws of perception. We interpreted gestalt-laws of homogeneity, similarity, proximity and figure and ground in link with color, spatial contrast at the level of regions and objects to produce feature contrast map. The model of top-down mechanism aims to use a formal computational model to describe the background connectivity of the attention and produce the priority map. Integrating both mechanisms and applying to salient object detection, our results have demonstrated that the proposed method consistently outperforms a number of existing unsupervised approaches on five challenging and complicated datasets in terms of higher precision and recall rates, AP (average precision) and AUC (area under curve) values.

Cognitive fusion of thermal and visible imagery for effective detection and tracking of pedestrians in videos

In this paper, we present an efficient framework to cognitively detect and track salient objects from videos. In general, colored visible image in red-green-blue (RGB) has better distinguishability in human visual perception, yet it suffers from the effect of illumination noise and shadows. On the contrary, the thermal image is less sensitive to these noise effects though its distinguishability varies according to environmental settings. To this end, cognitive fusion of these two modalities provides an effective solution to tackle this problem. First, a background model is extracted followed by a two-stage background subtraction for foreground detection in visible and thermal images. To deal with cases of occlusion or overlap, knowledge-based forward tracking and backward tracking are employed to identify separate objects even the foreground detection fails. To evaluate the proposed method, a publicly available color-thermal benchmark dataset Object Tracking and Classification in and Beyond the Visible Spectrum is employed here. For our foreground detection evaluation, objective and subjective analysis against several state-of-the-art methods have been done on our manually segmented ground truth. For our object tracking evaluation, comprehensive qualitative experiments have also been done on all video sequences. Promising results have shown that the proposed fusion-based approach can successfully detect and track multiple human objects in most scenes regardless of any light change or occlusion problem.

A novel hybrid algorithm of gravitational search algorithm with genetic algorithm for multi-level thresholding

Graphical abstractDisplay Omitted The multi-level thresholding is a popular method for image segmentation. However, the method is computationally expensive and suffers from premature convergence when level increases. To solve the two problems, this paper presents an advanced version of gravitational search algorithm (GSA), namely hybrid algorithm of GSA with genetic algorithm (GA) (GSA-GA) for multi-level thresholding. In GSA-GA, when premature convergence occurred, the roulette selection and discrete mutation operators of GA are introduced to diversify the population and escape from premature convergence. The introduction of these operators therefore promotes GSA-GA to perform faster and more accurate multi-level image thresholding. In this paper, two common criteria (1) entropy and (2) between-class variance were utilized as fitness functions. Experiments have been performed on six test images using various numbers of thresholds. The experimental results were compared with standard GSA and three state-of-art GSA variants. Comparison results showed that the GSA-GA produced superior or comparative segmentation accuracy in both entropy and between-class variance criteria. Moreover, the statistical significance test demonstrated that GSA-GA significantly reduce the computational complexity for all of the tested images.

A Hybrid Genetic Algorithm and Gravitational Search Algorithm for Global Optimization

The laws of gravity and mass interactions inspire the gravitational search algorithm (GSA), which nds optimal regions of complex search spaces through the interaction of individuals in a population of particles. Although GSA has proven effective in both science and engineering, it is still easy to suffer from premature convergence especially facing complex problems. In this paper, we pro- posed a new hybrid algorithm by integrating genetic algorithm (GA) and GSA (GA-GSA) to avoid premature convergence and to improve the search ability of GSA. In GA-GSA, crossover and mutation operators are introduced from GA to GSA for jumping out of the local optima. To demonstrate the search ability of the proposed GA-GSA, 23 complex benchmark test functions were employed, including unimodal and multimodal high-dimensional test functions as well as multimodal test functions with xed dimensions. Wilcoxon signed-rank tests were also utilized to execute statistical analysis of the results obtained by PSO, GSA, and GA-GSA. Experimental results demonstrated that the proposed algorithm is both efficient and effective.

A stability constrained adaptive alpha for gravitational search algorithm

Abstract Gravitational search algorithm (GSA), a recent meta-heuristic algorithm inspired by Newtons law of gravity and mass interactions, shows good performance in various optimization problems. In GSA, the gravitational constant attenuation factor alpha (α) plays a vital role in convergence and the balance between exploration and exploitation. However, in GSA and most of its variants, all agents share the same α value without considering their evolutionary states, which has inevitably caused the premature convergence and imbalance of exploration and exploitation. In order to alleviate these drawbacks, in this paper, we propose a new variant of GSA, namely stability constrained adaptive alpha for GSA (SCAA). In SCAA, each agents evolutionary state is estimated, which is then combined with the variation of the agents position and fitness feedback to adaptively adjust the value of α. Moreover, to preserve agents’ stable trajectories and improve convergence precision, a boundary constraint is derived from the stability conditions of GSA to restrict the value of α in each iteration. The performance of SCAA has been evaluated by comparing with the original GSA and four alpha adjusting algorithms on 13 conventional functions and 15 complex CEC2015 functions. The experimental results have demonstrated that SCAA has significantly better searching performance than its peers do.

Combinational Build-Up Index (CBI) for Effective Impervious Surface Mapping in Urban Areas

The distribution of urban impervious surface is a significant indicator of the degree of urbanization, as well as a major indicator of environmental quality. Hence, taking advantage of remotely sensed imagery to map impervious surface has become an important topic. Spectral indices have been developed due to its convenience to apply, among which feature extraction approach has shown superiority in reliability and applicability. However, impervious surface is often confused with bare soil when the current existing indices are used as well as their sensor-specific limitations. In this study, a new index, combinational build-up index (CBI), is proposed to extract impervious surface. The new index combines the first component of a principal component analysis (PC1), normalized difference water index (NDWI), and soil-adjusted vegetation index (SAVI), representing high albedo, low albedo, and vegetation, respectively, to reduce the original bands into three thematic-oriented features. The new index was tested using various remote sensing images at different spectral and spatial resolutions. Qualitative and quantitative assessments of the accuracy and separability of CBI, together with the comparison with other existing indices, were performed. The result of this study indicates that the proposed method is able to serve as an effective impervious index and can be applied widely.

Locally informed gravitational search algorithm

Gravitational search algorithm (GSA) has been successfully applied to many scientific and engineering applications in the past few years. In the original GSA and most of its variants, every agent learns from all the agents stored in the same elite group, namely Kbest. This type of learning strategy is in nature a fully-informed learning strategy, in which every agent has exactly the same global neighborhood topology structure. Obviously, the learning strategy overlooks the impact of environmental heterogeneity on individual behavior, which easily resulting in premature convergence and high runtime consuming. To tackle these problems, we take individual heterogeneity into account and propose a locally informed GSA (LIGSA) in this paper. To be specific, in LIGSA, each agent learns from its unique neighborhood formed by k local neighbors and the historically global best agent rather than from just the single Kbest elite group. Learning from the k local neighbors promotes LIGSA fully and quickly explores the search space as well as effectively prevents premature convergence while the guidance of global best agent can accelerate the convergence speed of LIGSA. The proposed LIGSA has been extensively evaluated on 30 CEC2014 benchmark functions with different dimensions. Experimental results reveal that LIGSA remarkably outperforms the compared algorithms in solution quality and convergence speed in general.

DMMOGSA: Diversity-enhanced and memory-based multi-objective gravitational search algorithm

Multi-objective optimization (MOO) is an important research topic in both science and engineering. This paper proposes a diversity-enhanced and memory-based multi-objective gravitational search algorithm (DMMOGSA). We combine the memory of the best states of individual particles and their population in their evolution paths and the gravitational rules to construct a new search strategy. Under this strategy, the position and mass states of each particle are updated based on the memory associated with it and the current states of all particles in the current population in terms of their gravitational forces on it. A novel diversity-enhancement mechanism is also employed to control the velocity of each particle for traveling to a new position. Experiments were conducted on 12 well-known benchmark functions, and for each function the results of DMMOGSA were compared with those of SPEA2, NSGA-II and MOPSO. Our results show that DMMOGSA can reduce the effect of premature convergence and achieve more reliable performance on most of the tested cases.

Joint bilateral filtering and spectral similarity-based sparse representation: a generic framework for effective feature extraction and data classification in hyperspectral imaging

Abstract Classification of hyperspectral images (HSI) has been a challenging problem under active investigation for years especially due to the extremely high data dimensionality and limited number of samples available for training. It is found that hyperspectral image classification can be generally improved only if the feature extraction technique and the classifier are both addressed. In this paper, a novel classification framework for hyperspectral images based on the joint bilateral filter and sparse representation classification (SRC) is proposed. By employing the first principal component as the guidance image for the joint bilateral filter, spatial features can be extracted with minimum edge blurring thus improving the quality of the band-to-band images. For this reason, the performance of the joint bilateral filter has shown better than that of the conventional bilateral filter in this work. In addition, the spectral similarity-based joint SRC (SS-JSRC) is proposed to overcome the weakness of the traditional JSRC method. By combining the joint bilateral filtering and SS-JSRC together, the superiority of the proposed classification framework is demonstrated with respect to several state-of-the-art spectral-spatial classification approaches commonly employed in the HSI community, with better classification accuracy and Kappa coefficient achieved.

Stratified spectral mixture analysis of medium resolution imagery for impervious surface mapping

Linear spectral mixture analysis (LSMA) is widely employed in impervious surface estimation, especially for estimating impervious surface abundance in medium spatial resolution images. However, it suffers from a difficulty in endmember selection due to within-class spectral variability and the variation in the number and the type of endmember classes contained from pixel to pixel, which may lead to over or under estimation of impervious surface. Stratification is considered as a promising process to address the problem. This paper presents a stratified spectral mixture analysis in spectral domain (Sp_SSMA) for impervious surface mapping. It categorizes the entire data into three groups based on the Combinational Build-up Index (CBI), the intensity component in the color space and the Normalized Difference Vegetation Index (NDVI) values. A suitable endmember model is developed for each group to accommodate the spectral variation from group to group. The unmixing into the associated subset (or full set) of endmembers in each group can make the unmixing adaptive to the types of endmember classes that each pixel actually contains. Results indicate that the Sp_SSMA method achieves a better performance than full-set-endmember SMA and prior-knowledge-based spectral mixture analysis (PKSMA) in terms of R, RMSE and SE.