SYSTEMS AND METHODS FOR SIGNAL CODING AND DECODING WITH ENTROPY-MINIMIZING KERNEL-BASED TRANSFORMATION
A system for coding and decoding of signals, such as pictures and video, intended for machine use is disclosed. An encoder implements a method for kernel-based transformation of the signal samples that reduces and preferably minimizes entropy of the input signal yet is equivalent to the input signal for the intended machine operation. As a result, the transformed signal sample can be compressed using fewer bits than the original. The encoded signal is received by a complimentary decoder and decoded for machine use.
1 . An encoder for signal coding with entropy-minimizing kernel-based transformation comprising:
a transformer module, the transformer module applying a set of transform kernels for a machine task and determining a transformed equivalent of the input signal featuring reduced entropy; and
a compressor module receiving the transformed signal and compressing the transformed signal to generate a compressed bitstream for transmission on a communication channel.
2 . The encoder of claim 1 , wherein the transform kernels are m×m matrices.
3 . The encoder of claim 2 , wherein the process of determining a transformed equivalent of the input signal further comprises:
unrolling each m×m transform matrix into a one-dimensional vector;
stacking the transform vectors into a single transform matrix;
partitioning the input signal into m×m signal patches;
unrolling each m×m signal patch into a one-dimensional signal vector;
using the transform matrix, determine a resultant vector, r, for each signal vector; and
for each resultant vector, determine the transformed signal as t=K T (KK T ) −1 r where K T is the transpose of the transform matrix.
4 . The encoder of claim 3 , wherein the input signal is a video signal.
5 . The encoder of claim 4 , wherein the transform kernels are Sobel kernels.
6 . The encoder of claim 4 , wherein the transform kernels are convolution kernels.
7 . The encoder of claim 4 , further comprising a machine model coupled to the transformer module and wherein the transform kernels are determined at least in part based on the machine model.
8 . The encoder of claim 4 , wherein the transformed signal is compressed by the compressor using a compression standard selected from the group including VVC, HEVC, and AV1.
9 . The encoder of claim 4 wherein the transformed signal is compressed by the compressor using a compression standard optimized for video coding for machines.
10 . The encoder of claim 1 , wherein the input signal is a video signal.
11 . The encoder of claim 10 , wherein the transform kernels are Sobel kernels.
12 . The encoder of claim 10 , wherein the transform kernels are convolution kernels.
13 . The encoder of claim 10 , further comprising a machine model coupled to the transformer module and wherein the transform kernels are determined at least in part based on the machine model.
14 . A method for encoding a signal with reduced entropy comprising:
obtaining a set of X transform kernels for a machine task, the kernels each in the form of an m×m matrix;
unrolling each m×m transform matrix into a one-dimensional 1×(m×m) transform vector;
stacking the transform vectors into an X×(m×m) transform matrix;
partitioning the input signal into m×m signal patches, and for each signal patch:
unrolling each m×m signal patch into a one-dimensional signal vector;
using the transform matrix, determine a resultant vector, r, for each signal vector; and
for each resultant vector, determine the transformed signal as t=K T (KK T ) −1 r where K T is the transpose of the transform matrix; and
compressing the transformed signal to form a compressed bitstream.