TRDAN: Lightweight Dual-Stream Attention Network with FPGA Acceleration for Industrial Gas Leak Detection

Gas leak detection in complex industrial environments faces critical challenges such as low thermal contrast, high false-alarm rates, and real-time processing demands. To address these limitations, this study proposes a lightweight gas detection framework named Thermal-RGB Dual-stream Attention Network (TRDAN), which integrates infrared thermography with RGB visual context for FPGA-accelerated edge deployment. The architecture introduces three key innovations: (1) A dual-stream attention mechanism that enhances infrared-to-temperature variation associations, significantly improving detection accuracy and robustness; (2) An Adaptive Dynamic Feature Fusion (ADFF) module employing learnable gating coefficients to enable dynamic cross-modal interactions; (3) FPGA-implementation optimization achieving inference acceleration. Experimental results demonstrate TRDAN's superiority over state-of-the-art methods, with 23.6% higher accuracy (mAP) and 320 ns latency on Xilinx Zynq-7000. It maintains robust performance under complex environmental conditions, validated through rigorous ablation studies. The successful FPGA deployment confirms its industrial application potential, providing a high-precision, low-cost solution for edge inspection devices.