Hybrid DQN-Based Low-Computational Reinforcement Learning Object Detection With Adaptive Dynamic Reward Function and...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep reinforcement learning (DRL) for object detection reduces region proposals and computational overhead.
• Current DRL methods lack precision due to poor image state representation and unstable agent learning.

Purpose of the Study:

• To develop a DRL-based object detection method that improves precision and reduces computational cost.
• To address the limitations of existing DRL approaches in image state representation and agent learning stability.

Main Methods:

• Low-dimensional RepVGG (LDR) feature extractor for reduced memory and fitting difficulty.
• Hybrid DQN (HDQN) to improve agent's state-action determination in complex environments.
• Adaptive Dynamic Reward Function (ADR) for dynamic reward adjustment.
• ROI Align-based bounding box regression network (RABRNet) for enhanced localization precision.

Main Results:

• Achieved 74.4% mAP on VOC2007, 76.2% mAP on COCO2017, and 75.2% Precision on SF dataset.
• Demonstrated superior precision compared to advanced DRL methods.
• Exhibited significantly lower computational cost (1.43G FLOPs) than existing DRL and mainstream methods.

Conclusions:

• LHAR-RLD offers highly accurate object localization with minimal computational demands.
• The method is well-suited for applications on resource-constrained devices.
• This approach advances DRL-based object detection by balancing precision and efficiency.