Image denoising via adaptive dictionary learning based on single-pixel imaging

Abstract

Based on the principle of compressed sensing, a single-pixel imaging system (single-pixel camera) was built in our laboratory, and our single-pixel camera can successfully acquire the picture of imaging scene of a distant building. However, the process of single-pixel imaging is easily disturbed by noise. In order to reduce the interference of noise in the imaging results as much as possible, we apply the method of adaptive dictionary learning to denoise the reconstructed image in our single-pixel camera. The block-coordinate decent idea and K-SVD algorithm are used to complete the dictionary learning by training on the noisy reconstructed image in the method of block-by-block scanning, which doesn’t require the extra corpus of image patches, so the method of dictionary learning is self-adaptive to the imaging picture that need to be denoised. The overcomplete DWT (discrete wavelet transform) dictionary is applied as the known initial dictionary that will be trained in the adaptive dictionary learning. The effectiveness of the adaptive dictionary learning has been demonstrated by the simulation experiment and the practical imaging experiment, the method of adaptive dictionary learning based on the initialization of overcomplete DWT dictionary effectively reduces the imaging noise and improves the image reconstruction accuracy and imaging quality.