An efficient reparameterized small object detection transformer for thermal infrared images

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Small object detection in thermal infrared images is challenging due to low contrast and limited texture.
• Deployment on edge platforms like Unmanned Aerial Vehicles (UAVs) faces computational constraints.

Purpose of the Study:

• To develop an efficient Transformer-based framework for infrared small object detection.
• To address computational limitations for real-time deployment on resource-constrained platforms.

Main Methods:

• Proposed PWL-RTDETR framework featuring a Partial Convolutional Reparameterization Block (PConvRep-Block) for efficient computation.
• Introduced WTRCSPNeck, a lightweight neck architecture with CNCSPELAN and WTConv for enhanced multi-scale feature aggregation.
• Implemented Layer-Adaptive Magnitude-based Pruning for model compression and sparsification.

Main Results:

• PWL-RTDETR demonstrated superior accuracy compared to state-of-the-art models on HIT-UAV and LLVIP infrared datasets.
• Achieved significant reductions in model parameters and Floating Point Operations (FLOPs).
• The model proved suitable for real-time applications in resource-constrained environments.

Conclusions:

• PWL-RTDETR offers an effective solution for accurate and efficient small object detection in infrared imagery.
• The framework's efficiency and accuracy make it viable for real-time perception tasks on edge devices like UAVs.