Improved Niching Particle Swarm Optimization And Its Application In Image Retrieval

Particle Swarm Optimization (PSO) is involved in one of the simplest swarm intelligence optimization algorithms, which could be applied in numerous fields such as neural network training and fuzzy system controlling. In recent years, PSO has been used to solve multiple optimization problems. However, some parameters that are required by many popular niching methods are difficult to obtain on the basis of human limited knowledge, which prevents it from being widely used in practical problems.A multimodal function optimization strategy depending on niching PSO is proposed, which does not require any parameters. This algorithm located particles' local optima depending on two aspects, the ratio of a particle's fitness and swarm's fitness and the Euclidean-distance among them. Alternatively, the radius of each niche is based on the distance between the local optima and other common swarms that regard that swarm as local optima. In the experimental analysis and comparison for several groups of functions that emphasizes different testing aspects, the algorithm proposed in this paper is superior to the traditional algorithms such as SPSO and FER-PSO, in terms of success rate and the converging speed, which shows the excellent performance of the algorithm.In the field of content-based image retrieval, single visual feature is usually difficult to reflect the whole content of images, therefore image retrieval based on features combination is becoming a hot research focus. An algorithm is initially proposed, which combines local color histogram and Scale Invariant Feature Transform and concentrates on the spatial distribution of color and affine invariance at the same time. Alternatively, two strategies are proposed to give these two features suitable weights, one is based on the user's relevant feedback, the other is dependent on the proposed niching PSO algorithm. In the experiment, the performance of image retrieval based on features combination is superior to that based on single feature, and the proposed algorithms that contributes to distribute due weights to relevant features are also superior to the fixed weight strategy.