151 results about "Vector quantization" patented technology

A method is provided for mapping the latent space of a Vector Quantized Variational AutoEncoder (VQ-VAE) to polynomial basis vectors. The method includes training a VQ-VAE model on a dataset to obtain a set of codebook vectors representing the latent space; defining a polynomial basis for the latent space, the polynomial basis containing terms up to a predetermined order; mapping each codebook vector to the polynomial basis by determining polynomial coefficients that represent each codebook vector in terms of the polynomial basis; and using the polynomial coefficients to reconstruct and manipulate latent space representations.

The invention discloses a method and a system for generating 3D (three-dimensional) human motion under text driving by utilizing a 2D (two-dimensional) video. The method comprises the following steps of: acquiring the video and preprocessing to obtain a two-dimensional key point sequence and text description; the two-dimensional key point sequence passes through a spatiotemporal feature adapter to obtain a potential spatiotemporal feature sequence, a residual vector quantizer quantizes and outputs a three-dimensional SMP L parameter sequence, and meanwhile, potential spatiotemporal features and a discrete Token sequence are mapped; preprocessing a text to extract a semantic vector, partially covering a Token sequence of a basic quantization layer, reconstructing a prediction sequence through a predictor in combination with the semantic vector, and obtaining a complete sequence through a refiner; constructing a total loss function and a text-to-action loss function to train the module; and inputting the text description and the basic quantization layer Token to a trained module, outputting a three-dimensional SMPL parameter sequence, and rendering to generate a three-dimensional human body grid and animation. According to the method, the end-to-end generation from the text to the three-dimensional SMPL action is realized only by two-dimensional key points and text description.

The invention relates to the technical field of wireless semantic communication, and discloses a semantic information transmission method, device and equipment and a storage medium, a sending end extracts features of to-be-transmitted data, and matches the features with vectors in a preset discrete vector codebook to obtain discrete feature vectors; and determining an orthogonal frequency division multiplexing time-frequency resource distribution grid according to the feature importance weights of the discrete feature vector and the initial feature vector, generating a sending signal, and transmitting the sending signal to a receiving end. And the receiving end inputs the semantic feature vector obtained after processing the received signal into the vector quantization variational self-decoder, and the semantic feature vector is matched with the vector in the discrete vector codebook again to correct the transmission error, thereby obtaining the reconstructed data to be transmitted. Discretization processing of the features is achieved through a discrete vector codebook, important semantic feature bits are quantized according to the importance weight of each feature and then distributed to orthogonal frequency division multiplexing for transmission, so that key semantic information is protected, and anti-interference transmission of the semantic information is achieved.

A method for training a model for generating a representation of long-term motion from a text input comprises: training a motion encoder of an autoencoder to compress and map an input motion into a latent representation comprising a sequence of latent vectors in a discrete latent space, each latent vector representing a fixed length of motion; training a quantization module to quantize the latent vectors to a sequence of quantized latent vectors in quantized latent space; and training a motion decoder to reconstruct the quantized sequence as a sequence of single-frame pose representations. A text encoder is trained to predict a latent sequence conditioned on a text input and a duration using the mapped latent representation as a target.

The invention provides a voice reconstruction method and system based on entropy coding residual quantization and frequency spectrum restoration, and relates to the technical field of artificial intelligence voice signal processing, and the method comprises the steps: obtaining an original voice waveform, inputting the voice waveform into a neural voice coding and decoding model, firstly entering a coder to map the input voice waveform into acoustic potential representation, and then entering a frequency spectrum restoration model; performing residual quantization on the acoustic potential characterization layer by layer through a residual vector quantization module, introducing a gating-based dynamic layer number selection mechanism and entropy regularization constraint, enabling bits to be adaptively distributed among different voice segments, reconstructing reconstructed acoustic features of the potential characterization, inputting the reconstructed acoustic features into a decoder, restoring the reconstructed acoustic features into a time domain waveform, and outputting the time domain waveform. And mapping to a logarithmic magnitude spectrum domain through a spectrum repairing module, predicting a residual error in the logarithmic magnitude spectrum domain and performing confidence gating fusion to obtain a complex spectrum, and outputting after time domain synthesis to obtain reconstructed speech. According to the invention, high fidelity, intelligibility and transmission reliability of the voice can be considered at an extremely low bit rate.

The invention discloses a high-fidelity three-dimensional Gaussian sputtering (3DGS) lightweight method for resource-constrained equipment. The method aims at solving the problems of high storage and computing resource consumption caused by the fact that a large number of parameters are stored in an existing 3DGS technology, and geometric distortion possibly occurring when details of a scene center are processed is overcome. The core of the method lies in a multi-stage progressive optimization framework, and the framework cooperatively applies four key technologies of Gaussian cutting and opacity regularization, dynamic spherical harmonic function adjustment, entropy constraint vector quantization and coordinate space shrinkage. Wherein in Gaussian clipping, redundant gauss are eliminated by quantifying the contribution degree of a Gaussian unit; the dynamic spherical harmonic function adjustment adaptively adjusts the order of color coding according to the scene complexity; the entropy constraint vector quantization is used for compressing a plurality of Gaussian attributes so as to realize more compact representation; and the coordinate space shrinkage is realized through nonlinear transformation, so that the rendering precision of details of the center of the scene is remarkably improved. According to the method, while the rendering precision and quality are kept, remarkable storage compression is realized, and the method is particularly suitable for deployment of resource-limited platforms such as mobile equipment.

The invention discloses a semantic-driven intelligent agent capability discovery method and device oriented to the intelligent agent Internet, and the method comprises the steps: firstly generating an intelligent agent structured portrait which covers the three-dimensional information of skills, roles and states; semantic coding is carried out on the portrait, and the portrait is converted into a high-dimensional semantic vector; then, the vector is partitioned, a codebook is generated in each subspace in a clustering mode, and the codebook comprises representative vectors and serial numbers of the representative vectors; and quantizing the sub-vectors into numbers based on a codebook, and splicing the numbers to form discrete identification codes of the agent index. And when the index is updated, incremental maintenance is executed according to the distance between the new agent sub-vector and the vector in the codebook. Meanwhile, a generative retrieval model is trained, task query is directly mapped into discrete identification codes, historical and new task samples are mixed for continuous learning in training, and stability constraints are introduced. According to the method, through semantic portraits, a quantitative index mechanism and memory enhancement continuous learning, an end-to-end retrieval and rapid updating capability discovery scheme is constructed.

The invention provides a semi-supervised signal modulation identification method based on multi-codebook discrete virtual adversarial training, and the method comprises the steps: carrying out the high-dimensional potential representation coding of input labeled and unlabeled sample signals through employing a vector quantization generative adversarial network, so as to obtain a continuous potential representation; mapping the continuous potential representation into a discrete vector sequence through a quantizer; virtual confrontation disturbance meeting the condition that the L2 norm does not exceed epsilon is applied to the unlabeled sample at the input end, and a disturbance sample is generated; respectively inputting a labeled sample, a label-free sample and a disturbance sample into a classifier, calculating cross entropy loss and KL divergence consistency loss, and jointly optimizing parameters of an encoder, a codebook and the classifier in a weighted sum form; therefore, the discrete potential representation is obtained through the vector quantization generative adversarial network, disturbance is constructed for the unlabeled sample at the input end, and joint optimization is performed in combination with the cross entropy and the KL divergence loss, so that the recognition accuracy and robustness under the conditions of few labels and low signal-to-noise ratio are improved.

The embodiment of the invention provides a CSI (Channel State Information) feedback method based on Transform and entropy constraint vector quantization. The method is applied to the technical field of wireless communication. The method comprises the following steps: receiving an input angle-time delay domain channel matrix, and performing feature extraction on the channel matrix through a convolutional layer to obtain a first feature tensor; the first feature tensor is input into a trained Transform encoder for analysis processing, and a latent vector is obtained; performing quantization processing on the latent vector through the trained vector quantization variational auto-encoder to obtain discretization representation of the latent vector; performing dequantization processing on the discretized representation of the latent vector to obtain a dequantized latent vector; performing feature extraction on the dequantized latent vector through a convolutional layer to generate an initial channel feature; the initial channel characteristics are input into a trained Transform decoder for analysis processing, and final channel characteristics are obtained; and the final channel feature is mapped to the target resolution to obtain a reconstructed channel matrix, so that the reconstruction precision of the channel matrix is improved.

The invention discloses a hidden space adversarial sample generation method and system based on multi-scale feature separation, and the method comprises the steps: employing a neural network quantization training method based on straight-through estimation, and training a hierarchical vector quantization variational auto-encoder; carrying out differentiable Haar wavelet transformation on the input image by adopting a wavelet packet transformation algorithm, decomposing the input image into a low-frequency component and a high-frequency component, and realizing multi-scale feature separation; inputting the high-frequency component into a hierarchical vector quantization variational auto-encoder, and extracting and quantizing global high-frequency features and local high-frequency detail features; in the potential space, a learnable disturbance variable is introduced, a potential vector after disturbance is constructed, and the potential vector is reconstructed into an adversarial sample through a decoder; and based on a preset disturbance target, carrying out iterative optimization on the disturbance vector until a confrontation sample which satisfies an attack success condition and is optimized in visual quality is generated. According to the method, a wavelet domain variational auto-encoder and a hidden space iterative attack algorithm are fused, and an adversarial sample with high fidelity and clear interpretation is generated.

The invention relates to a transferable vector quantization alignment method and device based on unsupervised domain adaptation, and the method comprises the steps: extracting the features of a source domain and a target domain of a source domain data set and a target domain data set through a feature extraction unit, and calculating an overall feature distribution loss function through searching a feature item closest to the features in a codebook; calculating a local alignment loss function through a bottleneck layer, performing classification through a classifier to obtain a source domain pseudo-label and a target domain pseudo-label, calculating a cross entropy classification loss function according to the source domain pseudo-label and a corresponding truth value label, performing normalization processing on the target domain pseudo-label, introducing mutual information to obtain a sample weight in each target domain data set, and obtaining a sample weight in each target domain data set; and calculating mutual information weighted maximization confusion matrix loss functions, and updating parameters in each module by using the loss functions until convergence to obtain a target classification architecture with cross-domain extraction, feature alignment and time sequence signal classification capabilities. By adopting the method, the identification performance of the label-free target domain can be improved.

The invention provides an EEG characterization method based on wavelet neural quantization training and semantic alignment, and the training method comprises the steps: S1, obtaining training data which comprises a plurality of EEG signal samples, and each sample comprises a plurality of EEG sub-signals; s2, using the training data to train a wavelet quantization nerve marker and a nerve codebook thereof: using a vector quantization encoder of the wavelet quantization nerve marker to carry out discrete wavelet transform on input samples, generating a feature block obtained by discrete wavelet transform and corresponding to each sub-signal of each sample, and extracting a potential representation of each feature block; acquiring a closest discrete vector matched from a plurality of discrete vectors in the neural codebook according to the potential representation, and reconstructing the EEG sub-signal corresponding to the feature block according to the discrete vector matched with each feature block by using a neural decoder based on inverse wavelet transform; and training and updating the parameters of the nerve marker and discrete vectors of the nerve codebook according to the first total loss function.

The invention discloses a multi-scale space-time autoregression movement track generation method, which comprises the following steps of: sequentially carrying out spatial scale representation and time scale representation on an original movement track to obtain a plurality of space-time representation sequences; performing residual vector quantization processing on each time-space representation sequence to obtain a discretized quantization index tag sequence corresponding to each time-space representation sequence; performing scale-by-scale autoregression generation on each quantization index tag sequence to obtain a discretized prediction quantization index tag sequence corresponding to each spatio-temporal representation sequence; and restoring the plurality of predictive quantization index tag sequences into a complete moving track. According to the embodiment of the invention, the moving track generation method has obvious advantages in the aspects of generation efficiency, track structure consistency, multi-scale modeling capability, individual behavior expression and the like.

This application provides an adaptive semantic joint source-channel coding method, system, electronic device, and storage medium. The method includes: Step S1: acquiring raw input data and real-time channel state information, wherein the real-time channel state information includes at least the signal-to-noise ratio (SNR); Step S2: using a lightweight semantic coding module to extract multi-scale semantic features from the raw input data, and incorporating the SNR embedding vector in a feature modulation manner during the coding process to obtain channel-adaptive semantic features; Step S3: using an SNR embedding and channel-adaptive attention module to adjust the attention weights of the semantic features to obtain attention features; Step S4: using a dynamic codebook generation and vector quantization module to convert the attention features into a discrete codebook index sequence; Step S5: using a joint source-channel coding module to modulate and map the codebook index sequence into complex channel symbols and transmit them.

The invention discloses an electric power vision large model multi-scale semi-supervised target detection method and system, and the method comprises the steps: training a multi-scale vision word segmentation device based on a vector quantization knowledge distillation framework, and constructing an electric power semantic enhanced multi-scale vision codebook, the multi-scale visual codebook comprises a top-layer codebook used for encoding a global structure of the power equipment and a bottom-layer codebook used for encoding defect local features of the equipment; based on the multi-scale visual codebook, adopting a mask image modeling task to pre-train a visual large model on an unmarked power image; performing semi-supervised fine tuning on the visual large model by using the labeled electric power image data and the unlabeled electric power image data; using an SMLS fine tuning strategy to update the visual large model parameters; and inputting a to-be-detected power image into the updated visual large model, and outputting a detection result of the multi-scale target. According to the method, the problems of scarcity of annotation data, model semantic gaps and high calculation complexity in an electric power scene in the prior art are solved.

An encoding device comprising: a quantization circuit that generates a quantization parameter that includes information about a vector quantization codebook; and a control circuit that sets the number of available bits according to conditions for encoding based on the difference between the number of bits available for encoding of the target sub-vector and the number of bits for the quantization parameter of the target sub-vector.

The invention belongs to the technical field of computer vision, and particularly relates to an image semantic representation-oriented collaborative distillation and quantitative compression method and an innovative joint training framework, which deeply couples two processes of vector quantization and knowledge distillation, realizes collaborative optimization through bidirectional information interaction, and can realize high compression ratio and high compression efficiency at the same time. And the quality of semantic representation output by the model is kept to the maximum extent.

This invention reduces the number of encoding bits in vector quantization. The encoding apparatus includes: a quantization circuit that generates quantization parameters, the quantization parameters including first information related to the codebook of the vector quantization and second information related to the code vectors contained in the codebook; and a control circuit that uses a second number of bits to control the encoding of the first information of a sub-vector, the second number of bits being the difference between the first number of bits available for encoding the sub-vector in vector quantization and the number of bits of the quantization parameters of the sub-vector.

This application provides a method, system, electronic device, and medium for caching codebooks based on vector quantization. The method includes: acquiring a codebook to be processed; sorting the codebook to be processed to obtain a target codebook index; obtaining an index boundary based on the target codebook index and a codebook cache; the codebook cache being a cache for storing the codebook to be processed; comparing the index boundary and the target codebook index to obtain a judgment result; and storing the codebook to be processed in the corresponding cache within the codebook cache based on the judgment result. This application improves memory performance and execution efficiency by placing codebook entries at different locations in the GPU's memory hierarchy based on the codebook's usage frequency, thus solving the problems of low efficiency in shared memory and global memory.

The invention discloses a vehicle-mounted audio equalization method and system based on residual vector quantization, and relates to the technical field of vehicle-mounted electronic equipment.The method comprises the steps that a white noise audio signal is played through a vehicle-mounted loudspeaker, a microphone is used for collecting the audio signal lost through in-vehicle propagation, and time-frequency characteristics are obtained through digital preprocessing; performing discretization processing on the target IIR filter coefficient through a residual vector quantization encoder to generate a plurality of layers of token sequences; inputting the time-frequency feature matrix into a convolutional neural network for feature extraction to obtain a low-dimensional feature vector, and through a multi-layer perceptron prediction head, respectively obtaining classification probability distribution of each layer of token and predicting a token sequence; and inputting the predicted token sequence into a residual vector quantization decoder, reconstructing the token sequence into an IIR filter coefficient, loading the IIR filter coefficient into a graphic equalizer, and performing frequency response compensation on an input audio signal. According to the method, the audio frequency response curve can be quickly and adaptively corrected, and the in-vehicle tone quality and the user hearing comfort are remarkably improved.

The invention discloses an edge computing device-oriented super-dimensional computing system with low power consumption and low resource consumption, which is characterized in that a preprocessing unit adopts a binary convolution kernel to carry out convolution operation on input image data and binarizes a convolution result to obtain binary data; the super-vector coding unit maps binary data corresponding to the training data and the reasoning data into a training super-vector and a reasoning super-vector; the training unit generates n class super-vectors according to label clustering of the training super-vectors; a dynamic retraining unit compares the similarity of the training super-vector and the class super-vector to obtain a most similar class label, when the most similar class label is inconsistent with a real label, a retraining updating process is started, and after retraining is completed, n class super-vectors are quantized and then input into a reasoning classification unit; and the reasoning and classifying unit adopts matrix multiplication to compare the similarity of the reasoning super-vector and the quantized class super-vector, and outputs a classifying result. The precision can be improved, the power consumption is reduced, and the hardware overhead is reduced.

The invention discloses a multi-dimensional complexity characterization method for infrared target identification in a complex ground scene, which belongs to the technical field of infrared target identification and tracking, and comprises the following steps: comparing the difference between the current feature of a target and the historical reference feature of the target, and quantifying feature degradation; carrying out feature vector extraction by adopting a module of a Zernike moment; target motion feature vectors are extracted, and target motion complexity is calculated; extracting a target historical reference feature and a target local background feature, quantifying a difference between the features, and calculating a similarity representation local background suspected degree; extracting feature vectors and motion vectors of the real target and the false motion target, and quantifying feature differences and inter-frame motion differences to represent the interference degree of the false motion target; and the target feature degradation degree and the like are fused into comprehensive identification tracking complexity. According to the method, the tracking difficulty of the current frame or sequence can be reflected, and a theoretical basis and decision support are provided for online switching of algorithm strategies, dynamic allocation of system resources and scientific evaluation of cross-scene algorithm performance.

The invention provides a high-fidelity low-illumination image enhancement method and system based on one-step diffusion, and the method comprises the steps: training a codebook based on a vector quantization generative adversarial network through a normal illumination image, obtaining a normal illumination codebook, and obtaining a high-fidelity low-illumination image through a latent variable refinement and alignment model; refining and aligning conditional latent variables of low illumination acquired by a VAE encoder of a pre-trained diffusion model to obtain refined conditional latent variables and aligned conditional latent variables, and quantifying the refined conditional latent variables and the aligned conditional latent variables to obtain quantized latent variables; taking the refined conditional latent variable as a diffusion starting point through a denoising network model based on one-step diffusion, taking the quantized latent variable as a vector for replacing text embedding, and obtaining an enhanced latent variable corresponding to the normal illumination image; and decoding to obtain a final enhancement result by using wavelet-based jump connection. According to the method, the generalization of complex and diversified real images is ensured while the low-illumination image is enhanced in a high-fidelity manner.

This invention relates to the field of natural language processing, and in particular provides a method and system for quantifying skincare efficacy vectors by integrating multidimensional constraints and semantic intent. The method includes receiving four-dimensional heterogeneous semantic input from users; constructing a multidimensional knowledge base, which includes a semantic weight association library, an intent-vector mapping library, and a constraint-correction matrix library; calculating a dynamic priority scalar for each skin problem by combining the influence of skincare motivation and skincare goals on various major skin problems; calling the intent-vector mapping library to obtain the primitive efficacy vectors corresponding to each skin problem, weighting the primitive efficacy vectors using the dynamic priority scalar, and aggregating them using a max-pooling algorithm to generate a baseline demand vector; calling the constraint-correction matrix library to obtain the corresponding correction matrix according to the constraints, performing nonlinear correction processing on the baseline demand vector, and outputting the final target efficacy representation vector. This invention achieves accurate, safe, and personalized quantification of skincare efficacy.

The invention belongs to the field of image processing, and provides a lightweight three-dimensional model optimization method and system for a mobile terminal, and the method comprises the steps: extracting a skeleton anchor point and a skeleton connecting line which represent the object shape connectivity; converting the local geometry and color of the original triangular mesh into a sparse Gaussian primitive set based on skeleton geometry; performing off-line vector quantization on the original texture of the geometric surface of the skeleton to generate a finite-scale texture token word list; training a lightweight mapping and decoding module at a mobile terminal side; loading and displaying an object skeleton, and determining a skeleton neighborhood needing to be loaded by taking the skeleton as a reference; carrying out placeholder display according to the appearance and thickness parameters of the Gaussian primitives; and calling a mobile terminal side mapping and decoding module to interpret the texture tokens into a candidate texture token set, retrieving the texture token with the highest similarity with the tokens in the candidate texture token set in a texture token word list, gradually replacing placeholder display, and recovering the original texture effect of the target patch.

This invention discloses a method for implementing bit allocation and quantization in audio compression, comprising the following steps: Step 1: Obtaining the audio signal, the perceptual importance of the audio signal, and the preset total target number of bits; Step 2: Based on the bit allocation vector corresponding to the minimum value of the calculated objective function, allocating bits to each frequency band of the audio signal, the sum of the bits allocated to each frequency band being the total target number of bits; Step 3: Feeding the audio signal and the bit allocation results for each frequency band into a residual vector quantization network for actual feature discretization, obtaining the quantized residual, thereby completing the quantization. This invention achieves adaptive bit allocation and quantization, thereby ensuring perceptually optimized encoding of audio and avoiding excessive differences in bit allocation between adjacent frequency bands that could lead to inter-band discontinuity artifacts in the reconstructed audio.

The application discloses a vector retrieval behavior simulation system, and the system comprises a nonvolatile memory simulation module, and the module comprises: a vector write model, which is used for performing conformity check on a base library vector and a write descriptor, quantizing the base library vector that passes the check, assigning a physical storage position to the base library vector according to a mapping strategy, writing the base library vector into a storage array, and generating a logical vector ID; a read model, which is used for locating a storage cell range in the storage array according to the physical storage position or the logical vector ID, reading base library quantization values, and restoring the base library quantization values into a data format recognizable by an upper layer according to format information in the write descriptor; and a query model, which is used for locating a storage cell range in the storage array according to each candidate vector in a candidate set, reading candidate base library quantization values stored in the storage cell range, processing a query vector, calculating a similarity index between the query quantization value and the candidate base library quantization values, and outputting a retrieval result based on the similarity index.

The invention discloses a multi-codebook semantic alignment vector quantization method for vision-language representation learning, and relates to the technical field of artificial intelligence image processing. The method comprises the following steps: encoding an attribute vocabulary into semantic vectors to form a global codebook and different attribute codebooks; constructing a multi-semantic perception coding module, and performing multi-scale feature extraction on the image to obtain global features and local features divided according to attributes; a multi-codebook vector quantization module is constructed, and the multi-codebook vector quantization module comprises a plurality of mutually independent codebook quantizers and is used for executing nearest neighbor search on continuous features and mapping the continuous features into discrete codeword indexes; constructing a cross-modal semantic alignment module, converting discrete tags into continuous vectors which can be processed by a large language model, and realizing alignment with language context in a semantic space; and designing a loss function, and training an optimization module by adopting a multi-task joint optimization strategy based on the loss function. According to the invention, high-fidelity image reconstruction and automatic image description generation with rich semantics and accurate details are realized.

The application discloses a QuickJS multi-interface collaborative optimization method and system of embedded equipment, relates to the technical field of JS multi-interface optimization, and comprises the following steps: S1, collecting interface task data, performing data preprocessing, identifying control difference and state change before and after the task; S2, setting a basic frame budget, constructing a task frame time consumption model, outputting a predicted frame time consumption, and comparing the current frame actual time consumption with the basic frame budget; S3, calculating the interface task priority, performing real-time sorting and scheduling, and establishing an asynchronous coroutine and task splitting and sticking mechanism; S4, outputting a structured rendering command, evaluating the redundancy of the control, and taking control adjustment measures; S5, constructing an error description vector, quantifying the deviation degree of the actual rendering result, and taking a rollback recovery measure; the problems that interface response depends on an event loop in the interaction between QuickJS and the underlying UI system, and that multiple JS interface tasks are simultaneously activated, resulting in abnormal interface rendering, are solved.

This invention discloses a high-quality hybridflow image compression method at extremely low bit depths, comprising: processing the input image by a continuous coding branch to obtain continuous bitstream bits, and decoding loopbacks to generate continuous reconstructed latent variables; processing the input image by a discrete branch to generate vector-quantized continuous latent variables; calculating a cross-stream difference map based on the continuous reconstructed latent variables and the vector-quantized continuous latent variables; performing masking decisions based on the cross-stream difference map to generate a final mask and transmission strategy map; organizing and encoding hierarchical syntax based on the final mask and transmission strategy map to obtain mask bits and index-related bits; and packetizing the continuous bitstream bits, mask bits, and index-related bits to output a bitstream. This invention solves the problems of inaccurate masking decisions and low indexing coding efficiency, achieving high fidelity and high perceptual quality at extremely low bit depths.