Tiny Generative Image Compression for Bandwidth-Constrained Sensor Applications

Abstract

Deep image compression algorithms based on Generative Adversarial Networks (GANs) are a promising direction to address the strict communication bandwidth limitations commonly encountered in IoT sensor networks (e.g. Low Power Wide Area Networks). However, current methods do not consider that the sensor nodes, which perform the image encoding, usually only offer very limited computation and memory capabilities, e.g. a resource-constrained tiny device such as a micro-controller. In this paper, we propose the first tiny generative image compression method specifically designed for image compression on micro-controllers. We base our encoder on the well-known MobileNetV2 network architecture, while keeping the decoder side fixed. To cope with the resulting asymmetric design of the compression pipeline, we investigate the impact of different training strategies (end-to-end, knowledge distillation) and integer quantization techniques (post-training, quantization-aware training) on the GAN-training stability. On the Cityscapes dataset, we achieve a compression performance that is very close to the state-of-the-art, while requiring 99% less SRAM size, 97% smaller flash storage and 87% less multiply-add operations. Our findings suggest that tiny generative image compression is particularly well suited for application-specific domains.