改进的Darknet噪声图像分类网络

A Noise Image Classification Network Based on Improved Darknet

针对现有噪声图像分类效率低的问题,提出一种改进的Darknet噪声图像分类算法。去掉Darknet网络输出部分的1×1卷积层,将第19层卷积核数量改为4,在网络最后加上Softmax层,实现网络分类功能。在网络passthrough层和第6~8层后分别引入Dropout层,在卷积层中引入L2正则化来避免网络过拟合。将网络第10层和第11层,第12层和第13层,第15层和第16层,第17层和第18层改为4个残差块,解决反向传播权值更新时梯度消失问题。从CIFAR-10数据集上取20 000张图片,经128×128尺寸变换后分别添加高斯噪声、泊松噪声、盐噪声和斑点噪声,对每张图片依类别进行One-hot编码,最后将图片和标签制作成训练集、验证集和测试集。4种算法实验结果对比表明,改进的Darknet网络对彩色噪声图像分类准确率可达0.904,远高于其他3种算法分类准确率。

Aiming at low efficiency of the existing noise image classification,a noise image classification algorithm based on the improved Darknet is proposed.The 1×1 convolution layer in the output part of Darknet network is removed,the number of convolution kernels in Layer 19 is changed to four,and Softmax layer is added to the end of the network,so that the classification function of the network is realized.Dropout is introduced after the passthrough layer and after Layer 6,Layer 7 and Layer 8 respectively,and L2 regularization is introduced in the convolution layer,so as to avoid network over-fitting.Layer 10 and 11,Layer 12 and 13,Layer 15 and 16,Layer 17 and 18 of the network are changed into four residual blocks to avoid gradient disappearance when updating the weights in back propagation.20 000 images are taken from CIFAR-10 data set,and four kinds of noise,that is,Gaussian,salt,speckle and Poisson,are added respectively after 128×128 size transformation.One-hot coding is carried out for each image according to its category.Finally, the images and labels are made into a training set,a verification set and a test set.The experimental results of the four algorithms show that the accuracy of the improved Darknet network for color noise image classification can reach 0.904,which is much higher than that of the other three algorithms.