Machine Learning in Action: Compressing Images with SVD

In the previous article, we learned about the principle of singular value decomposition (SVD) . Today, we will use the singular value decomposition of the matrix to compress the image.

Learn by doing

I made an online image compression application, you can feel it.

huggingface.co/spaces/beih…

The function is very simple, upload the image that needs to be compressed, select the compression ratio, and submit it.

Let's take a look at the implementation process

Compress images with SVD

The principle is simple:
decompose the image into RGB three-color matrices, perform singular value decomposition for each color matrix, and then select a specified number of features to compress the matrix.

core code

You can clone my huggingface for the complete code

huggingface.co/spaces/beih…

Core code 1:

p represents the percentage of singular values. According to the specified definition, the higher the definition, the lower the compression ratio, the more the number of singular values ​​extracted, and the less the picture will be distorted)

def rebuild_img(u, sigma, v, percent): 
    m = len(u)
    n = len(v)
    a = np.zeros((m, n))

    count = (int)(sum(sigma))
    curSum = 0
    k = 0
    while curSum <= count * percent:
        uk = u[:, k].reshape(m, 1)
        vk = v[k].reshape(1, n)
        a += sigma[k] * np.dot(uk, vk)
        curSum += sigma[k]
        k += 1
 
    a[a < 0] = 0
    a[a > 255] = 255
复制代码

Core code 2: It is mainly to define the implementation of the inderence function and the gradio front end

import os
os.system("pip install --upgrade pip")
os.system("pip install opencv-python-headless")
import cv2
import numpy as np
import gradio as gr
from func import rebuild_img

def inference(img,k):
    input_img = cv2.imread(img, cv2.IMREAD_COLOR)    
    u, sigma, v = np.linalg.svd(input_img[:, :, 0])
    R = rebuild_img(u, sigma, v, k)
    u, sigma, v = np.linalg.svd(input_img[:, :, 1])
    G = rebuild_img(u, sigma, v, k)
    u, sigma, v = np.linalg.svd(input_img[:, :, 2])
    B = rebuild_img(u, sigma, v, k)
    restored_img = np.stack((R, G, B), 2)
    return Image.fromarray(restored_img[:, :, ::-1])


gr.Interface(
    inference, 
    [
    gr.inputs.Image(type="filepath", label="Input"),gr.inputs.Slider(0, 1, 0.1,default=0.6,label= 'Compression ratio')], 
    gr.outputs.Image(type="pil", label="Output"),
    title=title,
    description=description,
    article=article
    ).launch(enable_queue=True,cache_examples=True,share=True)
复制代码

online

Gradio + Huggingface online machine learning application (purely free) I have introduced it many times, so I won't repeat it here, students who are not familiar with it, please move to my article: Tencent's algorithm, I moved it to the Internet, whatever Play!

Here are some minor problems and solutions.

Since cv2 is used, opencv-python needs to be installed, but the error is reported as follows:

  File "/home/user/.local/lib/python3.8/site-packages/cv2/__init__.py", line 8, in <module>
    from .cv2 import *
ImportError: libGL.so.1: cannot open shared object file: No such file or directory
复制代码

For this error, there are the following methods online:

1 yum install:

yum install libglvnd-glx
复制代码

2 Reinstall the opencv package:

pip uninstall opencv-python
pip install opencv-python-headless
复制代码

The first method requires root privileges. It is recommended to directly use the second method, which will save trouble.