Easz: An Agile Transformer-based Image Compression Framework for Resource-constrained IoTs

TL;DR

Easz tackles the practical challenge of neural image compression on resource-constrained edge devices by shifting the heavy encoding to a server while performing an edge-side erase-and-squeeze to produce a compact input for transmission. A row-based conditional sampler enables flexible, fine-grained compression levels, and a two-stage patchify with a lightweight transformer on the receiver reconstructs erased content with high perceptual quality. The approach maintains compatibility with existing codecs (e.g., JPEG/BPG) and demonstrates substantial reductions in edge computation, power, and memory usage, while achieving competitive or superior perceptual metrics on Kodak and CLIC datasets. In real-world edge-server testbeds, Easz delivers faster end-to-end performance and improved adaptability across compression levels, making neural image compression more viable for edgeIoT deployments.

Abstract

Neural image compression, necessary in various machine-to-machine communication scenarios, suffers from its heavy encode-decode structures and inflexibility in switching between different compression levels. Consequently, it raises significant challenges in applying the neural image compression to edge devices that are developed for powerful servers with high computational and storage capacities. We take a step to solve the challenges by proposing a new transformer-based edge-compute-free image coding framework called Easz. Easz shifts the computational overhead to the server, and hence avoids the heavy encoding and model switching overhead on the edge. Easz utilizes a patch-erase algorithm to selectively remove image contents using a conditional uniform-based sampler. The erased pixels are reconstructed on the receiver side through a transformer-based framework. To further reduce the computational overhead on the receiver, we then introduce a lightweight transformer-based reconstruction structure to reduce the reconstruction load on the receiver side. Extensive evaluations conducted on a real-world testbed demonstrate multiple advantages of Easz over existing compression approaches, in terms of adaptability to different compression levels, computational efficiency, and image reconstruction quality.

Figures (8)

• Figure 1: NN-based compressors face challenges on edge devices like the Jetson TX2, where loading and encoding an image can take over 10 seconds compared to a transmission latency of about 0.1 seconds.
• Figure 2: Left: Easz system overview. A image is going through a two-stage image patchify process, erase-and-squezee process. The squeezed image is then compressed using existing compressor and transmit. On the decompression stage, the erased patch is recovered. Right: (a) Proposed erase methods compared with random erase methods. T indicates an erased item in each row. (b) Different methods to re-organize un-erased image components.
• Figure 3: The proposed method outperforms random masking in terms of JPEG impact and reconstruction, resulting in a higher file saving ratio and lower MSE on Kodak dataset. The variable $p$ represents patch size.
• Figure 4: Easz preserves details better than SR methods via direct pixel prediction, improving PSNR and SSIM.
• Figure 5: Reconstruct process illustration.