融合STN和DenseNet的深度学习网络及其应用  

Deep Learning Network and Application Based on Fused STN and DenseNet

作　　者：张锡英[1] 宋宇鹏 陈曦 边继龙[1] ZHANG Xiying;SONG Yupeng;CHEN Xi;BIAN Jilong(School of Information and Computer Engineering, Northeast Forestry University, Harbin 150040, China)

机构地区：[1]东北林业大学信息与计算机工程学院

出　　处：《计算机工程与应用》2019年第17期131-136,共6页Computer Engineering and Applications

基　　金：中央高校基本科研业务费专项资金资助项目（No.2572016BB11）;国家自然科学基金（No.31500587）;黑龙江省科学基金（No.F2018002）

摘　　要：图像识别是计算机视觉的重要分支之一,具有重要的理论和实践意义。近年来,以深度卷积神经网络为代表的图像分类方法被成功地应用到各个领域。针对神经网络对输入数据敏感、训练时间长等问题,结合空间映射网络(Spatial Transform Network,STN)和密集神经网络(Dense Neural Network,DenseNet)两者的特性,提出一种新型网络结构ST-DenseNet。该网络能够对输入图片作不变性归一化处理,解决数据敏感问题的同时提高图像识别效果。在树种叶片公开数据集Leafsnap上实现了90.43%的识别准确率、87.75%的召回率和89.07%的F-Measure的实验结果,模型ST-DenseNet明显优于其他网络模型。Image recognition is one of the important branches of computer vision, which has important theoretical and practical significance. In recent years, image classification methods represented by deep convolutional neural networks have been successfully applied in various fields. In view of the sensitivity of neural networks to input data and the long training time, and combining the characteristics of Spatial Transform Network(STN)and Dense Neural Network(DenseNet), a new network structure, namely ST-DenseNet is proposed. The network solves the problem of data sensitivity and improves the effect of image recognition by way of normalized processing and immutability of the input image. On Leafsnap, the open data set of tree leaves, the experimental results of identification accuracy rate of 90.43%, recall of 87.75% and F-Measure of 89.07% are achieved. The model ST-DenseNet is obviously superior to other network models.

关 键 词：图像识别 深度卷积神经网络 空间映射网络 密集神经网络 树种识别 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]