Remote Sensing Scene Classification based on Generative Compressed-Domain

Abstract

Compressed-domain image classification refers to the direct feature extraction of image data in the form of compressed codestream to achieve scene classification tasks. In recent years, image compression based on deep learning has become an active research field, and its compressed codestream can retain rich high-level semantic features. Inspired by this, this paper proposes a deep representation model based on generative compressed-domain, and realizes the multi-scene classification of Remote Sensing Images. Specifically, the generative compression model optimizes training for image encoding, decoding and discrimination, so that the compressed bitstream can best retain or restore the image content. For the sparsely compressed bitstream, mapping high-dimensional feature to the input based on the residual structure, training the Soft-Max classifier for multi-scene classifcation. To a certain extent, the model is analogous to the decoder in the previous Generative Compression, imitating the process of reconstructing images to obtain deep features. Experiments show that the proposed method can obtain compressed-domain features with good discriminability and generalization, obtaining 94% classification accuracy results on the UCML scene classification data set.