基于双交叉熵的自适应残差卷积图像分类算法  

Adaptive residual convolution image classification algorithm withdual cross-entropy

作　　者：李伟 黄鹤鸣[1,2] LI Wei;HUANG He-ming(Computer College,Qinghai Normal University,Xining 810008,China;State Key Laboratory of Tibetan Intelligent Information Processing and Application,Qinghai Normal University,Xining 810008,China)

机构地区：[1]青海师范大学计算机学院,青海西宁810008 [2]青海师范大学藏语智能信息处理与应用国家重点实验室,青海西宁810008

出　　处：《计算机工程与设计》2023年第12期3670-3676,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(62066039、61662062)。

摘　　要：为弥补卷积神经网络在图像分类方面对颜色特征的不敏感,并生成更逼真的图像样本,提出一种基于双交叉熵的自适应残差卷积图像分类算法。将双交叉熵损失函数应用到深度卷积生成对抗网络中的判别模型;结合图像的主颜色特征和残差卷积神经网络提取的空间位置特征,运用改进的差分演化算法解决多特征融合权重系数的设定问题。实验结果表明,所提算法与传统的CNN算法相比,准确率明显提高10.75个百分点。双交叉熵损失函数可以提高判别模型区分生成图像与真实图像的能力,迫使生成模型生成更逼真的图像样本。To make up for the insensitivity of convolutional neural networks to color features in image classification and generate more realistic image samples,an adaptive residual convolution image classification algorithm with dual cross-entropy was proposed.The dual cross-entropy loss function was applied to deep convolutional generative adversarial networks to increase the diversity of samples.The main color features of images were effectively combined with the spatial position features extracted using residual convolutional neural networks.An improved differential evolution algorithm was used to optimize the weight value setting of multi-feature fusion.Experimental results show that the accuracy of the proposed algorithm is 10.75 percentage points higher than that of CNN.Besides,the dual cross-entropy loss function can improve discriminative model on distinguishing the generated images from the real images,thus forcing the generative model to generate more realistic image samples.

关 键 词：双交叉熵损失 生成对抗网络 卷积神经网络 多特征融合 自适应权重 改进的差分演化算法 图像分类 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]