基于差分进化优化的BP神经网络图像复原方法

An BP neural network image restoration method based on differential evolution optimization

针对back-propagating(BP)神经网络收敛速度慢、易陷入局部最小的问题,基于差分进化算法,改进其差分策略,提出随机缩放差分进化(random scaling-differential evolution,RSDE)优化的BP神经网络(RSDE-BP)图像复原方法.该方法用高斯噪声对无噪图像进行模糊处理,将模糊图像和原图像组成训练对,用于训练和优化RSDE-BP算法.最后利用训练好的BP神经网络对测试图像进行复原,从而达到去除噪声的目的.仿真结果表明,与BP神经网络、PSO-BP算法和DE-BP算法相比,所提出的算法收敛速度快,迭代次数少,且复原图像在峰值信噪比和结构相似性等指标方面有很好效果.与自适应全变差复原方法和二阶广义全变差正则项复原方法相比,该方法能够较好地恢复被噪声和模糊污染的图像,同时可以很好地保留图像的纹理和细节信息.

Aiming at the problems of slow convergence and local minimum of the back-propagating（ BP） neural network,an image restoration method based on random scaling-differential evolution（ RSDE） is proposed. In the proposed method,the noise-free images are blurred by the Gaussian noise. The blurred images and noise-free images are set to the training pairs,which are used to train and optimize the proposed RSDE-BP method. The optimized BP neural network is utilized to restore the testing images and remove the noises. The experimental results show that the convergence rate of the RSDE-BP algorithm is faster and the number of iterations is less than the traditional BP method,the PSO-BP method and the DE-BP method. In addition,the peak signal to noise ratio（ PSNR） and the structural similarity（ SSIM） are improved. Compared with the adaptive total variation（ ATV） image deblurring method and the blurred image restoration method based on the second-order total generalized variation（ TGV）regularization,the RSDE-BP method can effectively restore the images polluted by the noise and blur,while preserving the image texture and details more effectively.