Sign prediction for block-based video coding
Implementations of the disclosure provide a video decoding apparatus and method for transform coefficient sign prediction on a video decoder side. The method may include generating a plurality of candidate hypotheses for a set of candidate transform coefficients associated with a transform block of a video frame from a video. The method may further include selecting a hypothesis from the plurality of candidate hypotheses as a set of predicted signs for the set of candidate transform coefficients, wherein the hypothesis is selected based on a cost function calculated by extrapolating neighboring samples of the transform block in an extrapolation direction determined based on a dominant gradient direction. The method may also include estimating original signs for the set of candidate transform coefficients based on the set of predicted signs and a sequence of sign signaling bits received from a video encoder.
1 . A video decoding method for transform coefficient sign prediction, comprising:
generating, by one or more processors, a plurality of candidate hypotheses for a set of candidate transform coefficients associated with a transform block of a video frame from a video;
selecting, by the one or more processor, a hypothesis from the plurality of candidate hypotheses as a set of predicted signs for the set of candidate transform coefficients, wherein the hypothesis is selected based on a cost function calculated by extrapolating neighboring samples of the transform block in an extrapolation direction, wherein the extrapolation direction is determined based on a dominant gradient direction, wherein the dominant gradient direction is a direction with the largest magnitude in a histogram of gradient obtained based on an accumulated magnitude of gradients for each of a plurality of angular directions; and
estimating, by the one or more processors, original signs for the set of candidate transform coefficients based on the set of predicted signs and a sequence of sign signaling bits received from a video encoder.
2 . The video decoding method of claim 1 , further comprising:
receiving a bit stream comprising the sequence of sign signaling bits and quantized transform coefficients associated with the transform block;
generating dequantized transform coefficients from the quantized transform coefficients; and
updating the dequantized transform coefficients based on the estimated original signs for the set of candidate transform coefficients.
3 . The video decoding method of claim 2 , further comprising:
applying an inverse primary transform and an inverse low-frequency non-separable transform (LFNST) to the dequantized transform coefficients to generate residual samples in a residual block corresponding to the transform block.
4 . The video decoding method of claim 1 , further comprising:
determining the dominant gradient direction for extrapolating neighboring samples of the transform block.
5 . The video decoding method of claim 4 , wherein determining the dominant gradient direction further comprises:
selecting templates of neighboring samples for the transform block;
applying a gradient filter window to the templates to calculate gradients for the respective templates; and
selecting the dominant gradient direction based on the calculated gradients.
6 . The video decoding method of claim 5 , wherein selecting the dominant gradient direction based on the calculated gradients further comprises:
determining the histogram of gradient based on the calculated gradients, wherein each entry of the histogram of gradient corresponds to an accumulated magnitude of the gradients at a predefined angular direction; and
selecting the angular direction in the histogram of gradient that corresponds to the maximum accumulated magnitude as the dominant gradient direction.
7 . The video decoding method of claim 6 , wherein determining the histogram of gradient based on the calculated gradients further comprises:
calculating an angle and a magnitude of the gradient of each template;
converting the angle of the gradient to one of the predefined angular directions in the histogram of gradient; and
adding the magnitude of the gradient to the accumulated magnitude corresponding to the predefined angular direction converted from the angle of the gradient.
8 . The video decoding method of claim 5 , wherein the dominant gradient direction is determined to be the extrapolation direction for extrapolating the neighboring samples of the transform block in calculating the cost function, when a percentage of the neighboring samples of the transform block that belong to the dominant gradient direction exceeds a first predetermined threshold; or
the dominant gradient direction is determined to be the extrapolation direction for extrapolating the neighboring samples of the transform block in calculating the cost function, when a ratio between a magnitude of the gradient of the dominant gradient direction and a sum of magnitudes of all gradients exceeds a second predetermined threshold; or
the dominant gradient direction is determined to be the extrapolation direction for extrapolating the neighboring samples of the transform block in calculating the cost function, when a percentage of the neighboring samples of the transform block that belong to the dominant gradient direction exceeds a first predetermined threshold and that a ratio between a magnitude of the gradient of the dominant gradient direction and a sum of magnitudes of all gradients exceeds a second predetermined threshold.
9 . The video decoding method of claim 5 , wherein an orthogonal direction of the dominant gradient direction is determined to be the extrapolation direction for extrapolating the neighboring samples of the transform block in calculating the cost function.
10 . The video decoding method of claim 2 , wherein generating the dequantized transform coefficients from the quantized transform coefficients further comprises:
determining quantization indices of the dequantized transform coefficients based on transform coefficient levels of the quantized transform coefficients.
11 . The video decoding method of claim 10 , wherein determining quantization indices of the dequantized transform coefficients further comprises: for each quantized transform coefficient,
selecting a quantizer between two predefined scalar quantizers based on the transform coefficient level of a quantized transform coefficient preceding the quantized transform coefficient; and
determining the quantization index of the corresponding dequantized transform coefficient based on the selected quantizer applied to the quantized transform coefficient.
12 . The video decoding method of claim 11 , wherein the quantizer is selected between the two predefined scalar quantizers according to a state transition machine, wherein each state of the state transition machine corresponds to one of the two predefined scaler quantizers,
wherein the video decoding method further comprises:
determining a current state of a state transition machine based on a previous state of a state transition machine and a parity of the transform coefficient level of a quantized transform coefficient preceding the quantized transform coefficient; and
selecting the predefined scaler quantizer corresponding to the current state as the quantizer.
13 . The video decoding method of claim 10 , wherein:
the set of candidate transform coefficients are selected from the dequantized transform coefficients based on magnitudes of the quantization indices of the dequantized transform coefficients.
14 . The video decoding method of claim 10 , wherein the set of candidate transform coefficients are selected from the dequantized transform coefficients based on influence scores of the dequantized transform coefficients on reconstructed border samples of the transform block,
wherein the influence scores of the dequantized transform coefficients on the reconstructed border samples are measured as an L1 norm or an L2 norm of a variation of the quantization index of each dequantized transform coefficient on the reconstructed border samples.
15 . The video decoding method of claim 1 , wherein generating the plurality of candidate hypotheses for the set of candidate transform coefficients further comprises:
determining a plurality of combinations of sign candidates for the set of candidate transform coefficients based on a total number of candidate transform coefficients in the set of candidate transform coefficients; and
applying a template-based hypothesis generation scheme to generate the plurality of candidate hypotheses for the plurality of combinations of sign candidates, respectively.
16 . The video decoding method of claim 1 , wherein the sequence of sign signaling bits for the set of candidate transform coefficients indicate whether original signs of the candidate transform coefficients are identical to predicted signs of the candidate transform coefficients.
17 . The video decoding method of claim 16 , wherein the sequence of sign signaling bits includes a first bin with a value of zero,
wherein estimating the original signs for the set of candidate transform coefficients based on the set of predicted signs and the sequence of sign signaling bits further comprises:
adopting the predicted signs of a first group of candidate transform coefficients corresponding to the first bin in the sequence of sign signaling bits as the original signs of the first group of candidate transform coefficients.
18 . The video decoding method of claim 17 , wherein the sequence of sign signaling bits further includes a second bin with a value of one and a set of additional bins for informing corresponding correctness of the predicted signs of a second group of candidate transform coefficients,
wherein estimating the original signs for the set of candidate transform coefficients based on the set of predicted signs and the sequence of sign signaling bits further comprises:
correcting the predicted signs of the second group of candidate transform coefficients according to the additional set of bins before adopting the predicted signs as the original signs of the second group of candidate transform coefficients.
19 . A video decoding apparatus for transform coefficient sign prediction, comprising:
a memory configured to store a video comprising a plurality of video frames; and
one or more processors coupled to the memory and configured to:
generate a plurality of candidate hypotheses for a set of candidate transform coefficients associated with a transform block of a video frame from a video;
select a hypothesis from the plurality of candidate hypotheses as a set of predicted signs for the set of candidate transform coefficients, wherein the hypothesis is selected based on a cost function calculated by extrapolating neighboring samples of the transform block in an extrapolation direction, wherein the extrapolation direction is determined based on a dominant gradient direction, wherein the dominant gradient direction is a direction with the largest magnitude in a histogram of gradient obtained based on an accumulated magnitude of gradients for each of a plurality of angular directions; and
estimate original signs for the set of candidate transform coefficients based on the set of predicted signs and a sequence of sign signaling bits received from a video encoder.
20 . A non-transitory computer-readable storage medium having stored therein a bitstream comprising video information to be decoded by acts comprising:
generating a plurality of candidate hypotheses for a set of candidate transform coefficients associated with a transform block of a video frame from a video;
selecting a hypothesis from the plurality of candidate hypotheses as a set of predicted signs for the set of candidate transform coefficients, wherein the hypothesis is selected based on a cost function calculated by extrapolating neighboring samples of the transform block in an extrapolation direction, wherein the extrapolation direction is determined based on a dominant gradient direction, wherein the dominant gradient direction is a direction with the largest magnitude in a histogram of gradient obtained based on an accumulated magnitude of gradients for each of a plurality of angular directions; and
estimating original signs for the set of candidate transform coefficients based on the set of predicted signs and a sequence of sign signaling bits received through a bit stream from a video encoder,
wherein the bit stream is stored in the non-transitory computer-readable storage medium.