Title: Low-Latency Object Detection via Dual-Integrated Single-Lens Infrared Computational Imaging

Abstract: While computational imaging facilitates the development of compact infrared systems, deep-learning workflows that merge image reconstruction with object detection frequently suffer from significant inference latency. Current acceleration techniques typically focus on compressing the reconstruction network, often neglecting physical priors inherent to the optical path, which results in a compromise between processing speed and accuracy. To address this, we introduce the Physics-aware Dual-Integrated Network (PDI-Net), a framework designed for low-latency operation that combines infrared reconstruction with object detection by embedding optical priors directly into the learning mechanism. During training, PDI-Net employs a supervised U-Net; however, during inference, it utilizes a semi-U-Net encoder that shares features directly with a YOLO-based detector, thereby eliminating the need for complete image reconstruction. To reconcile the disparity between features optimized for reconstruction fidelity and those required for detection semantics, we developed the physics-aware large-small bridge (PALS-Bridge). This component leverages field-dependent point spread function priors to adaptively modulate multiscale convolutional branches. Additionally, a physics-informed optical degradation simulation pipeline was created for both training and validation purposes. The proposed method has been implemented on a single-lens infrared camera, achieving a system weight reduction of approximately 50% relative to conventional multi-lens configurations. In evaluations on the M3FD benchmark under low-SNR conditions, PDI-Net decreased inference time by 84.06% compared to a Rec+Det approach using pruning, while simultaneously increasing mAP@0.5:0.95 by 5.07%. These findings highlight the potential of compact, low-latency computational infrared imaging for enabling real-time object detection on resource-constrained devices.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC