DNN-Aware Image Compression via Gradient-Driven Quantization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing compression algorithms for IoT devices are designed for the human vision system (HVS) rather than deep neural networks (DNNs, leading to low compression ratios or high DNN inference errors due to limited wireless bandwidth and sensitivity to input noise.
Innovation Solution
GRACE (GRAdient-driven Compression for Edge) is a novel DNN-aware compression technique that analyzes a target DNN's perception model to generate an optimized compression strategy, reducing bandwidth usage without compromising inference accuracy by compressing components with low sensitivity in the DCT frequency domain.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If existing compression algorithms designed for human vision system are used, then compression ratio is improved, but DNN inference accuracy deteriorates due to sensitivity to input noise
Solution Approach 1:
The patent changes the compression parameters by deriving a DNN-specific quantization table from the target DNN's gradients in the DCT frequency domain, rather than using fixed HVS-based quantization tables. This allows the compression to preserve frequencies critical to DNN inference while discarding less important ones, resolving the contradiction between compression ratio and inference accuracy
Solution Approach 2:
The patent implements feedback by using the target DNN's gradient information to guide the compression process. The quantization table is derived from the DNN's perception model, creating a closed-loop system where the DNN's sensitivity to different frequencies informs the compression strategy, thereby maintaining inference accuracy while achieving compression
2Reliability
If high quality image and video are sent to edge for inference, then DNN inference accuracy is improved, but wireless bandwidth consumption increases causing scalability issues
Solution Approach 1:
The patent changes the quality parameter by optimizing the quantization table specifically for DNN inference requirements rather than using generic quality settings. This allows achieving adequate DNN inference accuracy with lower quality compression, thereby reducing bandwidth consumption and enabling scalable deployment
Solution Approach 2:
The patent extracts only the essential frequency components that are critical for DNN inference by using DNN-specific quantization tables. Non-essential frequency components are discarded, reducing the amount of data that needs to be transmitted over the wireless link while preserving the information necessary for accurate inference
3Device complexity
If DNN inference is offloaded to edge servers, then computing power requirements of IoT devices are reduced, but wireless bandwidth limitations hinder deployment scalability
Solution Approach 1:
The patent changes the compression parameter optimization from HVS-based to DNN-based, creating quantization tables that are specific to the target DNN's frequency sensitivity. This enables efficient compression tailored to the inference task, reducing bandwidth requirements and enabling scalable offloaded inference deployment
Solution Approach 2:
The patent performs preliminary action by pre-computing the DNN-specific quantization table offline based on the target DNN's gradient information. This pre-computed table is then used during real-time compression, enabling efficient bandwidth utilization without compromising inference accuracy during the actual deployment
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
GRACE achieves superior compression performance, reducing image size by 23% for semantic segmentation and bandwidth consumption by 90% for classification tasks while maintaining similar inference accuracy compared to existing methods like JPEG, with no additional overhead.
Implementation Method 1
convert the plurality of probe images from a spatial domain to a frequency domain
Implementation Method 2
use the quantization table that is unique to the target DNN to generate quantized coefficients
Data Source
AI summary
Example method includes: transmit a plurality of probe images from an Internet of Things (IoT) device at an edge network to a server hosting a target deep neural network (DNN), wherein the plurality of images are injected with a limited amount of noise; receive a feedback comprising a plurality of discrete cosine transform (DCT) coefficients from the server hosting the target DNN, wherein the plurality of DCT coefficients are unique to the target DNN; generate a quantization table based on the feedback received from the server hosting the target DNN; compress a set of real-time images using the generated quantization table by the IoT device at the edge network; and transmit the compressed set of real-time images to the server hosting the target DNN for DNN inferences.