zhangmozhe / microshift_compression

Microshift Compression: An Efficient Image Compression Algorithm for Hardware

Microshift: An Efficient Image Compression Algorithm for Hardware

This is the Matlab and Verilog implementation of the TCSVT paper "Microshift: An Efficient Image Compression Algorithm for Hardware"

Arxiv paper: https://arxiv.org/abs/2104.09820

IEEE link: https://ieeexplore.ieee.org/document/8529272

Github code: https://github.com/zhangmozhe/microshift_compression

Synthesis tutorial: https://github.com/zhangmozhe/microshift_compression/tree/master/VLSI%20design%20flow

Description

Microshift is a lossy image compression algorithm that can be efficiently implemented on Hardware with extremely low power consumption.

• When testing on dataset, it can compress images to 1.25 BPP with a resulting quality that outperforms state-of-the-art on-chip compression algorithms (PSNR=33.16, SSIM=0.90).
• An efficient VLSI architecture is proposed and is implemented on an FPGA.
• The results on the ASIC design further validate the low hardware complexity and high power efficiency.
• Our method is promising for low-power wireless vision sensor networks (WVSN).

Matlab code

Example of image compression and decompression

First, add the Toolkits folder and its subfolders to the searching path.

addpath(genpath('.Toolkits/'))

User can run the sample for compression and decompression by running:

run example_testbench_low_power.m

Note that this Matlab code is written in exactly the same style as the hardware implementation, and this causes long run-time. To speedup, we compile the compression file (compression_hdl_lowpower.m) into MEX file. We pre-compiled the MEX on windows, and users can compile the MEX file if using a different Matlab environment:

run compression_hdl_lowpower_coder_script.m

We provide a faster testbench that includes the MEX compression:

run example_testbench_low_power_faster.m

We provide two decompression modes as described in the paper:

• decompression_mode = 1: faster decompression

• decompression_mode = 2: MRF decompression, features better quality at the cost of longer run-time

Verilog generation

We can use HDLcoder to automatically generate the Verilog file from our compression Matlab file (compression_hdl_lowpower.m) by:

open HDL_generation.prj

Verilog code

We take use of the generated Verilog files and synthesize the design from it. We use the Global Foundry 180nm process. One can synthesize the design using Design Vision by running the following script:

dc_shell-t -f dc_new_216.tcl | tee log&

Note: make sure .synopsys_dc_setup is correctly set up so that the Design Vision can locate the foundry files.

Project structure

verilog
└───pre_synthesis
│   │   image.dat: image file
│   │   subimage*.dat: simulation outputs generated from Modelsim
│   │	*.v: Verilog design files; compression_fifo_hdl.v is the top design
└───post_synthesis: synthesized netlist from Design Vision
└───post_layout: post-layout netlist after routing in Encounter
└───script: contains the scripts for the Modelsim simulation
└───inout_images: testing images
└───matlab_files: contain the necessary MATLAB files for the project
└───image.dat: image binary file
└───ImageData_generation.m: matlab file that generates the image binary data

Pre-synthesis simulation

Directly run the Matlab script

run test_framework.m

• choose the image name: e.g. image_name = input_images/elaine.bmp
• choose the Modelsim simulation mode: set the GUI_mode variable
  • GUI mode: visualize the waveform in the Modelsim
  • command mode: this features faster Modelsim simulation

Post-synthesis simulation

Run the script test_framework.m for the joint simulation:

• choose the image name: e.g. image_name = input_images/elaine.bmp
• choose the Modelsim simulation mode by setting the GUI_mode variable
  • GUI mode: visualize the waveform
  • command mode: faster simulation

Explanation for the script test_framework.m

• the function ImageData_generation.m will generate the image.data files according to the chosen image content
• Modelsim simulation is run by calling the .do script in the Matlab
• after the simulation ends, the Modelsim will generate the compressed bitstream files subimage*.dat
• test_decompress is called to decompress the image from the bitstream files

You can simulate the synthesized netlist through through the following step:

• change to the directory post_synthesis/
• start the Modelsim in the terminal: vsim
• run the script: do ../script/my_testbench_post_synthsis.do, the simulation with the waveforms will automatically start

Explanation for the testbench file my_testbench_fpga_post_synthesis.v:

• set the clock_pedriod = 100ns
• back-annotate the SDF file to the component: u_compression_fifo

Explanation for the Modelsim script ./script/my_testbench_post_synthsis.do:

• Compile the library file csm18_neg.v to library work.target
• Compile the synthesized netlist compression_hdl_compiled.v to the library work.mapped
• note to add option into the .do script: -sdfnoerror +no_notifier

Explanation of the control signal in the Verilog design:

• clk = 5MHz

• RST_n = 0 for global reset;

• pixelin[7:0]: image raw data

• hstart = 1, whenever the first pixel of the horizontal line comes in;

• hend = 1, whenever the horizontal line ends

• vstart = 1, whenever meets the first pixel of the whole image

• vend = 1, whenever meets the last pixel of the whole image

• valid = 1, whenever the pixel is valid

More diagram and results

Illustration of the Microshift compression

Hardward implementation diagram

Image sensor with proposed compression ISP

Captured images with (left) and without (right) Microshift compression

Citation

@article{zhang2018microshift,
  title={Microshift: An Efficient Image Compression Algorithm for Hardware},
  author={Zhang, Bo and Sander, Pedro V and Tsui, Chi-Ying and Bermak, Amine},
  journal={IEEE Transactions on Circuits and Systems for Video Technology},
  volume={29},
  number={11},
  pages={3430--3443},
  year={2018},
  publisher={IEEE}
}