Remote Sensing Image Retrieval with Deep Features Encoding of Inception V4 and Largevis Dimensionality Reduction

摘要

Remote sensing image retrieval is an effective means to manage and share massiveremote sensing image data. In this paper, a remote sensing image retrieval methodhas been proposed, which adopts Inception V4 as the backbone network to extractthe deep features. To represent the low-level visual information of the remote sensingimage, the feature maps generated from the first Reduction Block of InceptionV4 through using 5 × 5 convolutional kernels are extracted and reorganized. Next,VLAD (Vector Locally Aggregated Descriptors) is exploited to encode the reorganizedfeatures to obtain a compact feature representation vector. The vector is cascadedwith the features extracted from the fully connected layers to form the overallfeature vector of the image. In order to avoid the problem of “Curse of Dimensionality”,Largevis dimensionality reduction method is utilized to reduce the dimensionalityof the image feature vector, while improving its discriminative capability.The dimensionality reduced feature vector is utilized for image retrieval with L2 distancemeasurement metric. Experimental results on the datasets of RS19, UCM andRSSCN7 have demonstrated that, compared with the existing methods, the proposedmethod can obtain state-of-the-art retrieval performance.