基于卷积神经网络的航标图像同态滤波去雾

Homomorphic Filtering for Navigation-Mark Image Dehazing with Convolutional Neural Network

为提高同态滤波算法对航标图像去雾的性能,采用改进的巴特沃斯滤波器作为同态滤波的传递函数。根据计算去雾后图像的标准差、平均梯度、亮度和二维熵等图像质量评价指标,分析图像锐化系数对图像去雾的作用效果。选取公开数据集ISCAS 2012作为卷积神经网络的输入训练数据,选取计算得到的使采用同态滤波去雾后的图像与真实图像的峰值信噪比最大的锐化系数作为输出数据,得到能计算最优锐化系数的卷积神经网络权重值。该方法能对测试样本数据输出相应的同态滤波器最优锐化系数。试验结果表明巴特沃斯同态滤波对航标图像去雾是有效的。

An improved Butterworth function is introduced into the homomorphic filtering algorithm with the transfer function to improve the image dehazing performance for navigation-mark images.The standard deviation,average gradient,brightness and two-dimensional entropy of the dehazed images are calculated.The effects of image sharpening coefficient are analyzed.In training for optimal weights of convolutional neural network,the data set ISCAS 2012 is input to the convolutional neural network to get an optimal sharpening coefficient,with which the ratio of s/n between the dehazed image and the original image reaches maximum.The design with Butterworth homomorphic filtering is verified through sample experiments.