The Unreasonable Effectiveness of JPEG: A Signal Processing Approach

Caption notes

Reducible follows JPEG from RGB pixel storage through encoding and decoding. Separating brightness from color makes chroma subsampling possible, drawing on differences in human visual sensitivity. He then treats image blocks as signals and uses the discrete cosine transform to express pixels as frequency coefficients. Dot products, orthogonality and the inverse transform explain why this change of representation is reversible. In many natural images, large coefficients concentrate at lower frequencies. Quantization can record less sensitive detail more coarsely, and run-length and Huffman coding then exploit repeated values. The central distinction is what the transform, quantization and coding each contribute, including where information is discarded. The reading follows the spoken explanation. Specific image comparisons and formulas should be checked in the video.