Image Compression: Traditional Schema and Machine-learning based schema

Feb. 12th 2020

Traditional Schema:

1. JPEG compression algorithm

Encoding Scheme

Step 1:

Input image, Convert it from RGB format to YCbCr format.

Step 2:

Split the image into blocks. 8X8 or 16X16.

Step 3:

Doing DCT transform to each block. The DCT matrix can be 8X8 or 16X16, which is relevant to your block size.

Step 4:

Make the pixel values after DCT transform be 0-centered. Values are subtracting 127.

Step 5:

Quantization step.  There is the quantization table.

QF, quality factor is being controled here. 2-255. bigger QF, much compression!!!

Step 6:

Huffman Coding to store the data. Order values and caluate the code.

 

Decoding Scheme:

 

Machine-learning based schema

1. Autoencoder

Using neural network layers to work as encoder and decoder. Minimize the MSE between input image and recovered image!

Evluate the image by PSNR value.

 

To be continued!