Infrared object detection faces persistent challenges: weak target-background contrast, low signal-to-noise ratio, and scarce texture. The study team observed that these conditions particularly undermine small-target detection and that conventional visible-light detectors may produce suboptimal features when applied directly to thermal imagery. To address that, the authors built TCLNet, a lightweight detector based on a YOLO-style backbone that explicitly models relative thermal contrast rather than attempting to recover physical temperature or radiance. The network aims to strengthen local, contrast-sensitive target cues and to limit background interference while remaining compact in parameter count.
The authors described the Thermal Contrast Encoder (TCE) as a thermally inspired feature-enhancement module. First, TCE projects multi-channel feature maps into a normalized single-channel response through successive convolutions, normalization, and nonlinearity. Next, the projected response is smoothed via average pooling plus a convolution to estimate a local background response that provides contextual reference. A normalized relative-contrast map is then computed from the projected response and its local background, with clipping and numerical-stability safeguards applied as reported by the source.
TCE also extracts spatial response gradients using fixed Sobel operators to capture local boundaries and small-target spatial structure. The module concatenates the original features, the relative-contrast map, and the aggregated gradient response; a learned fusion produces a modulation map that multiplies the original features, yielding enhanced outputs. The authors emphasized that TCE yields learned contrast-sensitive features and does not attempt to output calibrated temperatures, emissivity estimates, or other physical thermal parameters.
To improve interaction between contrast-aware and semantic representations, the paper presents the Thermal-Guided Cross-Feature Attention (TGCA) mechanism. In the reported design, contrast-enhanced features supply the Query while the semantic branch supplies Key and Value. The attention computation is performed within local windows to reduce computational cost and to limit attention to spatially relevant regions rather than distant background areas. The authors reported that this cross-feature, local attention helps preserve and amplify small-target signals that may otherwise be diminished by downsampling or dominated by clutter.
TCE and TGCA are integrated into a lightweight feature-extraction block called Thermal-Attentive Fusion C2f (TAF-C2f). The fusion block is intended to jointly optimize contrast-sensitive enhancement, local cross-attention, and semantic feature extraction while keeping the model parameter footprint small. The authors report an overall model size of approximately two million parameters, noting that the architecture was constructed to balance detection performance with deployment feasibility on resource-constrained devices.
Experiments were performed independently on three public datasets: M3FD, Teledyne FLIR ADAS v2 (denoted FLIR_2), and NUDT-SIRST1. The reported mean average precision at 50% IoU (mAP50) values for TCLNet were 60.5% on M3FD, 45.4% on FLIR_2, and 91.4% on NUDT-SIRST1. The authors compared these numbers to a YOLO12n baseline and reported absolute improvements of 12.8, 11.3, and 7.5 percentage points respectively. The paper states that these gains are especially evident in low-SNR and small-target conditions, and that the model still maintains a compact parameter count of roughly 2M.
The source clarified that no new primary image datasets were created; all experiments used publicly available third-party data and the paper provides pointers to the official repositories and portals for dataset access.
The authors explicitly stated that TCE does not recover absolute temperature, calibrated radiance, emissivity, or thermal-physical parameters; its outputs are learned contrast-sensitive representations evaluated by detection metrics. The original paper includes comparative experiments, ablation studies, visual analyses, and robustness evaluation; specific implementation hyperparameters and full experimental protocols are contained in the source and should be consulted for reproduction. The study notes avenues for future research in the paper’s final sections, and the authors discuss possible directions rather than definitive next steps. No competing interests or external funding were declared in the source.