A Deep learning accelerator for lossless image compression using TVM/VTA stack

Abstract

Image compression is a requirement for storing and transmitting images. Many fields like digital photography demand lossless image compression as it’s a requirement to reconstruct compressed images without a loss. Deep learning has opened so much room for improvements in image processing-related tasks. There are lossless image compression algorithms which use deep learning to achieve impressive compression ratio values. However, the time efficiency of lossless image compression might be affected due to using deep learning. Low-cost edge computing devices cannot use GPUs to accelerate deep learning algorithms. Deep Learning Accelerators (DLA) are the most feasible solution to eliminate time efficiency issues of deep learning-based algorithms for edge computing devices. Deep learning-based lossless image compression solution can be implemented in a System on Chip (SoC) with a Field Programmable Gate Array (FPGA). We propose a lossless image compression system in which a properly trained deep Convolutional Neural Network (CNN) is used to predict residual errors of LOCO-I-based pixel value prediction. Adaptive Arithmetic coding is applied to further improve the compression ratio. The main contribution of our approach is implementing the trained deep CNN in hardware using TVM/VTA stack. Finally, our proposed solution implements an end-to-end lossless image compression system by carrying out the prediction of residual error values of LOCO-I-based pixel value prediction in a Pynq-Z1 board (FPGA) while performing the rest of the tasks in a Python application. TVM/VTA stack stands as a bridge between the Python application and the FPGA. The proposed method yields a better compression performance with respect to state-of-the-art codecs according to the experimental results. The hardware implementation improves the time efficiency significantly enabling utilising the predictive power of deep CNNs for image compression systems. This is the first time a DLA is used effectively in a lossless image compression system, to the best of our knowledge