Simplification of Deep Neural Network-Based Object Detector for Real-Time Edge Computing.

Area of Science:

• Computer Vision
• Machine Learning
• Embedded Systems

Background:

• Deep neural networks (DNNs) offer powerful object detection capabilities but are computationally intensive.
• Deploying complex DNNs on resource-constrained edge devices requires significant model optimization.

Purpose of the Study:

• To develop and evaluate methods for simplifying and quantizing DNN-based object detectors for real-time edge device deployment.
• To identify optimal network simplification and quantization strategies for embedded systems.

Main Methods:

• Compared five channel pruning methods for residual blocks in DNNs, evaluating detection performance, parameter count, complexity, and speed.
• Assessed post-training quantization (PTQ) and quantization-aware training (QAT) on datasets with varying detection difficulty.
• Implemented and tested the optimized DNN on a Qualcomm QCS605 System-on-Chip (SoC).

Main Results:

• Identified the most effective channel pruning method for the specific edge device architecture.
• Determined that QAT is superior to PTQ for difficult object detection tasks on edge devices.
• Achieved real-time performance exceeding 10 frames per second on the target edge device.

Conclusions:

• The proposed method successfully enables the deployment of DNN-based object detectors on edge devices.
• The optimized models provide efficient real-time object detection suitable for embedded applications.