Object Detection Model Compression for Resource-Limited Devices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI-based object detection models require substantial computing resources, making them unsuitable for deployment on low-power, resource-limited devices, which often operate in environments without access to power grids and rely on batteries or energy harvesting, limiting their effectiveness in real-time video-based processing.
Innovation Solution
A method involving backbone replacement, image size reduction, model pruning, and quantization is employed to compress AI-based object detection models, enabling their deployment on resource-limited devices while maintaining detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If AI-based object detection models are deployed on resource-limited devices, then real-time processing capability is improved, but computing resource requirements exceed device capabilities
Solution Approach 1:
The patent segments the object detection model into multiple components (backbone network, detection head, etc.) and applies different compression techniques to each segment. The backbone network undergoes pruning and quantization, while the detection head is optimized separately, allowing the model to run on resource-limited devices while maintaining real-time processing capability.
Solution Approach 2:
The patent changes critical parameters of the model including precision (from 32-bit floating point to 8-bit integer through quantization), network depth (through pruning), and architecture (through backbone replacement with lighter alternatives). These parameter changes reduce computational requirements while preserving detection accuracy for real-time operation.
2Device complexity
If model compression techniques are applied, then device resource requirements are reduced, but detection model accuracy deteriorates
Solution Approach 1:
The patent applies preliminary quantization and pruning during the model training phase rather than as a post-processing step. By performing these compression techniques beforehand with proper regularization and retraining, the model learns compressed representations that maintain accuracy while reducing size for deployment on resource-limited devices.
Solution Approach 2:
The patent employs feedback mechanisms during compression where detection accuracy is continuously monitored and used to guide the pruning and quantization process. Important weights and filters are preserved based on their contribution to accuracy, while redundant elements are removed, achieving a balance between model size and detection performance.
3Measurement precision
If full-precision models are used, then detection accuracy is maintained, but energy consumption increases beyond battery capacity
Solution Approach 1:
The patent replaces expensive 32-bit floating point arithmetic operations with cheaper 8-bit integer operations through quantization. This substitution dramatically reduces energy consumption per operation, allowing the model to run continuously on battery-powered devices without requiring high-capacity batteries or frequent recharging.
Data Source
AI summary
Disclosed herein is a method for efficiently reducing the computational footprint of any AI-based object detection model, so as to enable its real-time deployment on computing resource-limited (i.e., low-power, embedded) devices. The disclosed method provides a step-by-step framework using an optimized combination of compression techniques to effectively compress any given AI-based object detection model.