Research On Noise Suppression Technologies Of Bayer Image Sequence

Bayer image is the raw output of CMOS or CCD sensors. It is unavoidably influenced by noise in low illumination. Noise not only affects the visual perceptual quality but also decreases compression ratio, increases the transmission bandwidth, and influences the efficiency of the subsequent intelligent analysis algorithms. Denoising on Bayer image can suppress noise from the source for the noise model complex problem caused by nonlinear transformation operation during image signal processing pipeline, such as demosaicking, color correction, and Gamma correction. This paper focuses on noise suppression of Bayer image sequence, especially toward the application for the case with stationary background. Our research is based on the change compensation. At first, detect the change area in the noisy image sequence. Then, do the temporal denoising on the unchanged background area and do the spatial denoising on the changed area. With this strategy, this dissertation does the several researches as follows:(1) Noise suppression on Bayer image based on luminance guided filter. Guided filter can unite the spatial relationship among each color component, which makes all red, green, blue color components have the same gradient. As for guided information selection, luminance component is chosen, which has the higher Signal-to-Noise-Ratio(SNR), compared to other color component. In order to evaluate the hardware realizability and real-time capability, the algorithm is implemented on FPGA platform. The experiment is carried out on both simulated noisy image sequence and real-captured noisy image sequence. It demonstrates that our algorithm is superior to other window based filter kernels. Bayer image noise removal is the spatial denoising part of image sequence spatiotemporal denoising method. It can be used as the single image denoising method too.(2) Real-time Bayer image sequence noise suppression based on luminance component downsampling. The real-time application on hardware devices is very important in daily life. This paper proposes a denoising algorithm based on Bayer luminance component downsampling change detection. The downsampling can effectively suppress noise influence on change detection and it is quite convenient for FPGA implementation. The temporal denoising takes the weighted averaging on the past denoised Bayer frames and the current one. The spatial denoising takes the method of(1). This algorithm is also implemented on FPGA to evaluate its real working performance. The experiment on simulated noisy image sequence and real-captured image sequence demonstrates that with small-scale noise, our algorithm is inferior to other state-of-the-art methods, such as VBM3 D, but with large-scale noise, its performance surpasses all of them. Since the application in low illumination with large-scale noise is the main work of denoising function, our algorithm has its actual engineering significance for digital imaging devices.(3) Bayer image sequence noise suppression based on shearlet representation. Besides the real-time application on hardware, the higher denoising ability is more important for post-processing task. This paper proposes the change detection method based on shearlet representation so as to improve the accuracy of the change area. In the background area, the Kalman temporal denoising is employed to obtain more excellent performance. In the change area, non-local means method is utilized as the prefilter on the luminance component, which can improve the SNR of the guided signal. The comparison experiment shows with small-scale noise, our algorithm is competitive with other state-of-the-art methods, and with large-scale noise, our algorithm has the dominant superiority, particularly in the background area.(4) Bayer image sequence noise suppression based on small motion vector characteristic. High-frame-rate image sequence has the advantage of crisp and fluid imagery. However, it also brings the noise problem for the short exposure time. This paper presents the containment change detection method according to small motion vector characteristic of high- frame-rate image sequence. This method further improves the change detection capability compared to the shearlet-based one. The temporal and spatial denoising methods are the same as(3). The comparison experiment shows for the image sequence which has the small motion vector characteristic, our algorithm is competitive with other state-of-the-art methods with small-scale noise, and has the higher performance with large-scale noise. Besides high-frame-rate image sequence, this algorithm is fit for all the image sequences having the small motion vector characteristic, such as the normal-frame-rate capturing for the slow motion object.