基于复合余弦基神经网络的图像去噪滤波器

Compound-cosine-basis neural network filter for removing image noise

针对图像受噪声干扰问题,提出了一种基于复合余弦基神经网络的图像去噪低通滤波器,以及保证该神经网络算法收敛的收敛定理,并给出了证明.该方法通过复合余弦基神经网络对幅频响应训练,调整网络权值,控制总的能量函数在给定的极小范围内,使得实际滤波器特性逼近理想图像去噪低通滤波器特性,从而达到满意的去噪效果.仿真结果显示,复合余弦基神经网络图像去噪滤波器各项特性接近理想滤波器.与传统的图像滤波器去噪方法比较,该方法设计的图像去噪滤波器具有更好的去噪效果.

A 2-D lowpass filter based on the compound-cosine-basis neural network（CNNL） was proposed to overcome the defects of conventional methods for removing image noise, and the convergence theorem which ensures that this algorithm is convergent was presented. This method manages to have the characteristics of the real filters and approach the characteristics of the ideal lowpass filters for removing image noise through the training of the compound-cosine-basis neural network, which adjusts the weight of the network along with the training, so that the energy function is less than the infinitesimal, then good resuits of removing image noise could be attained. The simulation results reached nearly ideal filter characteristics, and the performance of removing image noise using this filter was compared with the median filter. The experimental results showed that the effect of this method for removing image noise is better.