Image Encoder Quantization Step Adjustment for Bit Rate Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image encoding methods based on deep neural networks face challenges in efficiently adjusting bit rates without cumbersome retraining processes, as modifying the Lagrange multiplier λ requires retraining the network for each desired bit rate.
Innovation Solution
The proposed solution involves a training device and method that acquires latent variables and uses a cost function related to deviations between input and restored image data to train image encoders and decoders, allowing for quick adjustment of bit rates by modifying the quantization step Q, rather than retraining the network for different bit rates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If the Lagrange multiplier λ is modified to adjust bit rate, then the bit rate can be controlled, but the network must be retrained for each desired bit rate which increases time consumption and reduces efficiency
Solution Approach 1:
The patent introduces a preliminary quantization step that converts continuous latent variables into discrete quantized latent variables before entropy encoding. This quantization operation is performed once during encoding, and the quantized values are then used for entropy coding. By separating the quantization parameter Q from the network training process, the system allows bit rate adjustment through parameter modification rather than network retraining, thus saving time while maintaining adaptability.
Solution Approach 2:
The patent changes the parameter being adjusted from the Lagrange multiplier λ (which requires network retraining) to the quantization step Q (which does not require retraining). The cost function is modified to include a quantization term, and the training process optimizes the network to work with quantized latent variables. This parameter substitution allows bit rate control through simple parameter change rather than complex retraining, resolving the contradiction between adaptability and time consumption.
2Reliability
If conventional image encoding methods are used, then the encoding process is simple, but the performance in terms of bit rate-distortion tradeoff is inferior compared to deep neural network-based methods
Solution Approach 1:
The patent introduces quantized latent variables as an intermediary between the continuous latent variables from the encoder and the entropy coding process. The quantization step Q acts as a mediator that controls the bit rate while the neural network handles the complex transformation. This intermediary approach allows the system to benefit from both the performance of deep learning and the simplicity of conventional entropy coding, resolving the contradiction between performance and complexity.
Solution Approach 2:
The patent segments the encoding process into distinct stages: (1) encoding to obtain continuous latent variables, (2) quantization to obtain discrete quantized latent variables, and (3) entropy coding of the quantized values. This segmentation allows each component to be optimized independently - the neural network for performance and the quantization-entropy coding pipeline for efficiency and simplicity, thus resolving the contradiction between performance and complexity.
3Adaptability or versatility
If multiple networks are trained with different Lagrange multipliers for different bit rates, then various bit rates can be achieved, but the training process becomes cumbersome and inefficient
Solution Approach 1:
The patent designs a universal encoder network that can operate at multiple bit rates through a single training process. The network is trained with a cost function that includes quantization, and during operation, different bit rates are achieved by simply changing the quantization parameter Q rather than using different networks or retraining. This universal design provides multi-functionality while maintaining training simplicity, resolving the contradiction between adaptability and ease of manufacture.
Data Source
AI summary
Embodiments of this disclosure provide a training method, an image encoding method, an image decoding method and apparatuses thereof. The image encoding apparatus includes: an image encoder configured to encode input image data to obtain a latent variable; a quantizer configured to perform quantizing processing on the latent variable according to a quantization step to generate a quantized latent variable; and an entropy encoder configured to perform entropy coding on the quantized latent variable by using an entropy model to form a bit stream.