Systems and methods for image compression at multiple, different bitrates
Systems and methods for predicting a target set of pixels are disclosed. In one embodiment, a method may include obtaining target content. The target content may include a target set of pixels to be predicted. The method may also include convolving the target set of pixels to generate an estimated set of pixels. The method may include matching a second set of pixels in the target content to the target set of pixels. The second set of pixels may be within a distance from the target set of pixels. The method may include refining the estimated set of pixels to generate a refined set of pixels using a second set of pixels in the target content.
1. A computer-implemented method for encoding target content, the method being implemented in a computer system that comprises a non-transient electronic storage and one or more physical computer processors, comprising:
obtaining, from the non-transient electronic storage, target content and a corresponding quantization step size;
obtaining, from the non-transient electronic storage, a trained compression model, the trained compression model conditioned by training an initial compression model using training data, wherein the training data comprises (i) a set of training content, (ii) for each training content included in the set of training content, a plurality of quantization step sizes for quantizing the training content, and (iii) for each training content included in the set of training content, a plurality of reconstruction losses based on differences between the training content and a corresponding quantized training content that was quantized using a given quantization step size included in the plurality of quantization step sizes; and
applying, with the one or more physical computer processors, the trained compression model to the target content to generate compressed target content.
2. The computer-implemented method of claim 1 , wherein applying the trained compression model comprises:
encoding, with the one or more physical computer processors, the target content to generate encoded target content;
quantizing, with the one or more physical computer processors, the encoded target content using the corresponding quantization step size to generate quantized target content;
applying, with the one or more physical computer processors, a bottleneck to the quantized target content to generate decodable target content; and
decoding, with the one or more physical computer processors, the decodable target content to generate the compressed target content.
3. The computer-implemented method of claim 2 , wherein applying the bottleneck to the quantized target content comprises modifying the quantized target content such that
y
˜
=
y
+
U
(
-
1
2
,
1
2
)
·
s
where {tilde over (y)} represents the decodable target content, y represents the quantized target content, U represents uniform noise, and s represents the corresponding quantization step size.
4. The computer-implemented method of claim 1 , wherein training the initial compression model comprises:
a) applying, with the one or more physical computer processors, a first subset of the set of training content using a first quantization step size,
b) applying, with the one or more physical computer processors, a second subset of the set of training content using the first quantization step size and a second quantization step size;
c) repeating, with the one or more physical computer processors, b) by adding additional quantization step sizes until a maximum quantization step size is reached.
5. The computer-implemented method of claim 1 , wherein the corresponding quantization step size is concatenated with the target content in order to apply the trained compression model.
6. The computer-implemented method of claim 1 , wherein the target content comprises one or more of an image, a video frame, or a residual corresponding to the target content.
7. The computer-implemented method of claim 2 , wherein quantizing the encoded target content is defined by
Q
(
y
,
x
)
=
(
y
s
)
where y represents the encoded target content, s represents the corresponding quantization step size, and round(⋅) represents a rounding function that maps values to a closest integer.
8. The computer-implemented method of claim 1 , wherein the reconstruction losses are defined by
( x,x ′)= ( x,x ′)+γ·ε( {tilde over (y)} )
where (⋅,⋅(represents a rate-distortion loss value, (⋅,⋅) represents a reconstruction loss value based on given target content and corresponding quantized target content, γ represents a weight, ε(⋅) represents an entropy value, and represents the quantized target content.
9. A computer-implemented method for training a compression model that improves quantization of target content, the method being implemented in a computer system, the computer system including one or more physical computer processors and a non-transitory electronic storage, comprising:
obtaining, from the non-transient electronic storage, training data, the training data comprising (i) a set of training content, (ii) for each training content included in the set of training content, a plurality of quantization step sizes for quantizing the training content, and (iii) for each training content included in the set of training content, a plurality of reconstruction losses based on differences between the training content and a corresponding quantized training content that was quantized using a given quantization step size included in the plurality of quantization step sizes;
obtaining, from the non-transient electronic storage, an initial compression model;
generating a trained compression model by training, on the one or more physical computer processors using the training data, the initial compression model; and
storing, in the non-transient electronic storage, the trained compression model.
10. The computer-implemented method of claim 9 , further comprising:
obtaining, from the non-transient electronic storage, target content and a corresponding quantization step size; and
applying, with the one or more physical computer processors, the trained compression model to the target content and the corresponding quantization step size to generate compressed target content.
11. The computer-implemented method of claim 10 , wherein applying the trained compression model comprises:
encoding, with the one or more physical computer processors, the target content to generate encoded target content;
quantizing, with the one or more physical computer processors, the encoded target content using the corresponding quantization step size to generate quantized target content;
applying, with the one or more physical computer processors, a bottleneck to the quantized target content to generate decodable target content; and
decoding, with the one or more physical computer processors, the decodable target content to generate the compressed target content.
12. The computer-implemented method of claim 10 , further comprising:
encoding, with the one or more physical computer processors, a residual corresponding to the target content.