Signal-dependent Noise Suppression Method For CMOS Image Sensor Based On Stochastic Resonance Theory

The image sensor is the core component of the imaging system,and its performance directly determines the quality of the imaging.CMOS image sensor and CCD image sensor are the two most widely used sensors nowadays.Compared to CCD image sensor,CMOS image sensor have less power consumption,lower cost and higher levels of integration.However,in terms of noise,CMOS image sensor is not as good as CCD image sensor.Therefore,it is particularly important to study the noise suppression method of CMOS imaging devices.The traditional image denoising method is to filter out noise as much as possible,and with the advent of stochastic resonance(SR)theory,it is confirmed that the noise is not completely harmful,and it can be utilized to enhance the signal.Therefore,the innovation work of this paper is to design image denoising methods in three ways based on stochastic resonance under the CMOS image sensor signal-dependent noise model aiming at the lack of weakening the signal while filtering noise in the traditional image denoising algorithm.(1)From the perspective of signal energy and power,the effective noise intensity of signal-dependent noise and signal-independent noise is derived and substituted into the signal-to-noise ratio(SNR)formula of adiabatic approximation theory(AAT).The stochastic resonance phenomenon under the noise model is proved.when processing an image,in order to satisfy the requirement that the input signal must be a one-dimensional small signal according to the stochastic resonance theory,the noisy image is first normalized and then converted to a one-dimensional sequence by dimensionality reduction scanning.Input the one-dimensional sequence into the bistable system and solve the equation by Runge Kutta(RK)algorithm.The one-dimensional output sequence is restored to the two-dimensional image by inverse scanning and gray stretch transformation.The experimental results show that stochastic resonance occurs among the noise,image and bistable system.A part of the noise energy is transferred to the signal which enhances the image signal and achieves the effect of noise reduction.(2)Aimming at the characteristic of signal-dependent noise,that is,different gray values are affected by different noise intensities,this paper proposes a stochastic resonance image denoising algorithm based on image segmentation.Each block after image segmentation is processed by dimensionality reduction and then input into a bistable system.The results show that the image denoising effect is better when the system parameters of each image block are consistent.(3)In order to achieve a more ideal noise reduction effect,in this paper,we study the stochastic resonance system of cascade structure.To perform stochastic resonance processing on an image from the row direction and the column direction in turn and analyze its superiority in image denoising performance from both subjective and objective perspectives compared with the one-time stochastic resonance and stochastic resonance based on image segmentaton.On this basis,a lot of experiments have been performed by cascaded stochastic resonance image denoising algorithm and the traditional image denosing algorithms.The experimental results show that under low noise the Peak Signal to Noise Ratio(PSNR)after cascaded stochastic resonance image denoising is not as good as the traditional image denoising algorithms.However,as the noise intensity increases,the cascaded stochastic resonance gradually shows its good noise reduction performance.In terms of SSIM,regardless of low noise or strong noise,the cascaded stochastic resonance exhibit better results than other algorithms.