注意力机制融合深度神经网络的室内场景识别方法  

Method for Recognizing Indoor Scene Classification Based on Fusion Deep Neural Network with Attention Mechanism

作　　者：薛永杰 巨志勇[1] XUE Yong-jie;JU Zhi-yong(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海200093

出　　处：《小型微型计算机系统》2021年第5期1022-1028,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(81101116)资助.

摘　　要：针对现有室内场景识别方法仅通过关注视觉信息本身,而缺乏考虑图像中所含语义内容,提出一种基于长短期记忆神经网络和卷积神经网络的融合深度神经网络室内场景识别方法.首先使用labelImg工具为Visual Genome数据集图像生成位置描述符,经数据预处理算法处理后通过GloVe模型得到词向量.然后引入带有L2正则化的小批量梯度下降算法训练模型,将注意力机制与融合深度神经网络模型结合实现对位置描述符进行特征提取.最后通过Softmax函数进行场景分类.该文方法在所选Visual Genome数据集上取得了97.2%的识别准确率,结果表明该方法相较于传统机器学习方法和单一CNN方法具有更优识别准确率,表明了该方法在室内场景识别领域的有效性和可行性.To tackle the issues that the existing indoor scenes recognition methods only focus on the visual information but lack of considering the semantic content in the image,this paper proposed an indoor scene recognition method based on a fusion deep neural network which combines long short-term memory and convolutional neural network.Firstly,using labelImg to generate position describer for the pictures in Visual Genome dataset,after being processed by data pre-processing algorithm the word vector is obtained by the GloVe method.Secondly,the model is trained by mini-batch gradient descent strategy with L2 regularization,then combined the attention augmentation mechanism and the proposed fusion deep neural networks to extract the feature of position describer.Finally,the scene is classified by the Softmax function.In this paper,the proposed method achieved 97.2%accuracy on the Visual Genome dataset.The results show that the proposed algorithm achieved higher accuracy compared to the traditional machine learning method and single CNN method,which proves the effectiveness and feasibility of this method in the field of indoor scene classification.

关 键 词：室内场景图像描述 注意力机制 卷积神经网络 长短期记忆网络 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]