NLM3D infrared video denoising method based on spatio-temporal domain feature matching

Infrared raw image noise distribution randomness and variety, these noises will not only affect the imaging quality of infrared images, but also mixed with the details of the target scene, affecting the target detection and identification performance. In order to get better imaging effect, it is necessary to denoise the raw data of infrared images. In this paper, a NLM3D infrared video denoising method based on feature matching in time and space domains is proposed for the shortcomings of traditional time domain and non-local mean (Non-Local Means, NLM) denoising algorithms. Firstly, the displacement matrix is calculated by the dense fast optical flow method, from which the displacement matrix locates the search domains of the front and rear frames, then the optimal block matching algorithm is utilized to establish the 3D filtering window, and finally the multi-frame recursive structure is used for the NLM-based filtering denoising. The algorithm in this paper utilizes the simplified fast optical flow method to establish reliable motion estimation, which makes full use of the structural similarity information in the spatio-temporal domain, and can effectively remove the inter-frame noise, and meanwhile utilizes the improved NLM filtering based on the Mahalanobis distance metric of the block similarity, which can effectively suppress the streak and Gaussian noise. The experimental results show that compared with the time domain denoising algorithm, the algorithm in this paper does not produce the phenomenon of tail dragging and ghosting in multi-frame motion scenarios, and in terms of noise evaluation, the peak signal-to-noise ratio, roughness, and uniform region standard deviation of the algorithm in the paper are better than other algorithms.