研究目的
To propose and compare multi-threshold image segmentation procedures based on nature-inspired multi-objective optimization algorithms for human silhouettes detection in video sequences.
研究成果
The multi-objective Black Hole Algorithm (MOBHA) provides better segmentation results compared to MOPSO and MOGSA for the tested images, generating more Pareto solutions and achieving lower region counts for similar RMSE values. However, thresholding-based segmentation fails when gray levels are close, indicating a need for alternative methods in such cases.
研究不足
The algorithms start with random initial positions, leading to variability in results across runs. The segmentation may not perform well when background and foreground gray levels are similar, requiring additional geometric criteria. The study is limited to specific test images from the KTH dataset and may not generalize to all scenarios.
1：Experimental Design and Method Selection:
The study adapts Particle Swarm Optimization (PSO), Black Hole Algorithm (BHA), and Gravitational Search Algorithm (GSA) for multi-objective optimization. The optimization criteria are Root Mean Square Error (RMSE) and the number of segmented regions, both to be minimized. The algorithms are implemented and tested on gray-level images from video sequences.
2：Sample Selection and Data Sources:
Test images are extracted from the KTH human activity video dataset, consisting of 120x160 pixel frames with 256 gray levels. Images include scenarios with uniform backgrounds and varying human silhouettes.
3：List of Experimental Equipment and Materials:
A computer system with C++ implementation and the OpenCV library for image manipulation and processing functions.
4：Experimental Procedures and Operational Workflow:
Algorithms are run with specific parameters (e.g., number of individuals=25, iterations=200). For each algorithm, multiple runs are performed with random initializations. The Pareto front of solutions is computed and evaluated based on RMSE and region count.
5：0). For each algorithm, multiple runs are performed with random initializations. The Pareto front of solutions is computed and evaluated based on RMSE and region count.
Data Analysis Methods:
5. Data Analysis Methods: Solutions are compared visually and through Pareto front analysis. The number of Pareto solutions and their distributions in terms of RMSE and region count are tabulated and graphed for comparison.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
