Combining KPCA and PSO for Pattern Denoising

KPCA based pattern denoising has been addressed. This method, based on machine learning, maps nonlinearly patterns in input space into a higher-dimensional feature space by kernel functions, then performs PCA in feature space to realize pattern denoising. The key difficulty for this method is to seek the pre-image or an approximate pre-image in input space corresponding to the pattern after denoising in feature space. This paper proposes to utilize particle swarm optimization (PSO) algorithms to find pre-images in input space. Some nearest training patterns from the pre-image are selected as the initial group of PSO, then PSO algorithm performs an iterative process to find the pre-image or a best approximate pre-image. Experimental results based on the USPS dataset show that our proposed method outperforms some traditional techniques. Additionally, the PSO-based method is straightforward to understand, and is also easy to realize.