应用于图像分割的卷积神经网络参数简化模型  

A Parameter Simplification Convolutional Neural Network Model for Image Segmentation

作　　者：孙双林 杨倩 张优敏 SUN Shuanglin;YANG Qian;ZHANG Youmin(School of Software,Chongqing Institute of Engineering,Chongqing 400056,China)

机构地区：[1]重庆工程学院软件学院,重庆400056

出　　处：《重庆理工大学学报（自然科学）》2021年第3期145-151,共7页Journal of Chongqing University of Technology：Natural Science

基　　金：国家自然科学基金面上项目“安全压缩感知及其在资源非对等环境下图像隐私保护中的应用研究”(61572089);重庆市教委科技计划重点基金项目(KJZD-K201801901);重庆市教委科学技术研究项目资助(KJ1717371)。

摘　　要：在图像处理领域卷积神经网络普遍存在模型复杂、参数众多等问题,使得方法应用不便、时间开销不可控。针对这一问题,提出了一种混合规模的深度卷积神经网络,该网络架构基于扩展的卷积方法,以便在不同图像规模上采集到关键特征,并通过将深度特征紧密相连,从而控制参数的数量。实验结果表明,新方法可以在限制模型参数前提下达到相当高的准确率,并能有效降低过拟合的风险。In recent years,convolutional neural networks have been successfully applied to many image-processing related areas.However,the complexity of the models and huge number of parameters makes it difficult to implement and lead to high time consumption.To solve above problem,this article proposed a mixed-scale convolutional neural network based on dilated convolution strategy in order to capture features at different image scales.It densely connects all feature maps with each other,thus controls the number of parameters.Experimental result reveals the novel method can achieve quite high accuracy by less parameters.Moreover,it can reduce the risk of overfitting the training data.

关 键 词：卷积神经网络 图像分类 参数规模 准确率 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]