A⁴RSIC: Attention-Augmented Adversarial Autoencoder for Remote Sensing Image Compression

Reconstruction of images from the recent deep learning based models for image compression tasks does not necessarily corroborate with human perception and lose relevant inherent structural details. To address this, in this research work, a novel deep attention-augmented adversarial autoencoder based lossy remote sensing image compression <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(\mathbf{A}^{4}\mathbf{RSIC})$</tex> approach is proposed that achieves superior compression ratio compared to the state-of-the-art methods. The <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{A}^{4}\mathbf{RSIC}$</tex> model embeds an encoder-generator duo to generate a compact latent vector representation of the input image, which is then used to reconstruct the original image with minimal loss. The proposed model utilizes a discriminator, trained to differentiate between real and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{A}^{4}\mathbf{RSIC}$</tex> generated images using an adversarial approach, ensuring the fidelity of the reconstructed images. Furthermore, the model integrates a convolutional block attention module (CBAM) at relevant strategic points within the encoder and generator to enhance the salient features by leveraging spatial and channel-wise attention mechanism. This improves the quality of feature representation and compression performance. Experimental results demonstrate the efficacy of the proposed method compared to JPEG as well as other state-of-the-art methods.