CAJ | 학술논문

将堆叠去噪编码器应用于图像识别中，并针对自然图像检索领域，使用卷积自动编码器改进堆叠去噪编码器，采用非监督贪婪逐层训练算法初始化网络权重，使用反向传播算法优化网络参数。利用改进的堆叠去噪编码器提取自然图像特征，使用softmax分类器分类。最后将提取的特征结合尺度不变特征转换技术实现图像检索。实验证明改进型堆叠去噪编码器大幅降低网络训练时长的同时增强网络的容错能力，提高分类器的分类精度，最终改善图像的检索效果。
장퇴첩거조편마기응용우도상식별중，병침대자연도상검색영역，사용권적자동편마기개진퇴첩거조편마기，채용비감독탐람축층훈련산법초시화망락권중，사용반향전파산법우화망락삼수。이용개진적퇴첩거조편마기제취자연도상특정，사용softmax분류기분류。최후장제취적특정결합척도불변특정전환기술실현도상검색。실험증명개진형퇴첩거조편마기대폭강저망락훈련시장적동시증강망락적용착능력，제고분류기적분류정도，최종개선도상적검색효과。
Stacked denoising auto-encoder( SDA) is introduced into the image recognition. Convolutional auto-encoder ( CAE) is used to improve SDA in the area of natural image retrieval. The unsupervised greedy layer-wise training algorithm is used to initialize the weight of the network. The parameters of the network are optimized by the back propagation algorithm. The improved SDA is trained for extracting features from natural images and the softmax classifier is used for classification. Finally, the extracted feature combined with scale invariant feature transform ( SIFT) is used for realizing images retrieval. The experimental results show that the improved stacked denoising auto-encoder( ISDA) method can greatly reduce the time of network training, enhance the fault-tolerant ability of network, raise the classification precision of the classifier and eventually improve the image retrieval performance.