---
title: "TCLNet: Thermal contrast-aware lightweight network improves infrared object detection"
id: "plos-one-6-tclnet-a-thermal-contrast-aware-lightweight-network-for-infrared-object"
canonical_url: "https://medichelpline.com/clinical-feed/plos-one-6-tclnet-a-thermal-contrast-aware-lightweight-network-for-infrared-object"
content_type: "clinical_feed_article"
specialty: "Radiology"
source_name: "PLOS ONE (Medicine)"
source_url: "https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114"
published_at: "2026-10-22T14:00:00.000Z"
evidence_level: "Journal Feed"
license: "CC-BY-NC-4.0 / Informational Use"
---
# TCLNet: Thermal contrast-aware lightweight network improves infrared object detection
## Provenance & Clinical Metadata
- **Canonical URL:** https://medichelpline.com/clinical-feed/plos-one-6-tclnet-a-thermal-contrast-aware-lightweight-network-for-infrared-object
- **Specialty:** [Radiology](https://medichelpline.com/clinical-feed/radiology.md)
- **Primary Source:** PLOS ONE (Medicine)
- **Source URL:** [Original Journal Publication](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114)
- **Published At:** 2026-10-22T14:00:00.000Z
- **Evidence Rating:** Journal Feed
## Executive GIST (TL;DR)
- The authors propose **TCLNet**, a lightweight YOLO-based architecture tailored for infrared object detection, designed to balance detection accuracy with a small parameter budget. - A novel **Thermal Contrast Encoder (TCE)** learns a normalized single-channel response, a local background reference via average pooling and convolution, a normalized relative-contrast map, and Sobel-gradient features to produce a contrast-sensitive modulation map for feature enhancement. - The model adds a **Thermal-Guided Cross-Feature Attention (TGCA)** mechanism to mediate interactions between contrast-enhanced features (as Query) and semantic features (as Key/Value) with a local-window strategy to limit distant background interference and reduce computation. - A fused block, Thermal-Attentive Fusion C2f (TAF-C2f), integrates TCE and TGCA into a compact feature-extraction module, aiming to jointly optimize contrast enhancement, local cross-attention, and semantic extraction within a small parameter budget. - Experiments used three public datasets: M3FD, FLIR ADAS v2 (FLIR_2), and NUDT-SIRST1; reported mAP50 values were 60.5%, 45.4%, and 91.4% respectively, improving over a YOLO12n baseline by 12.8, 11.3, and 7.5 percentage points. - The model size is approximately 2 million parameters, which the authors report as maintaining compactness while improving detection, notably for low-SNR and small-target cases. - The study clarifies that TCE does not recover calibrated temperature, radiance, emissivity, or physical heat-diffusion parameters; outputs are learned contrast-sensitive features. - All experiments used publicly available third-party datasets; no new primary image dataset was generated. The source lists dataset access instructions and repositories for M3FD, FLIR ADAS v2, and SIRST. - The paper provides comparative experiments, ablation studies, visual analyses, and robustness evaluation; specific methodological hyperparameters, training schedules, and some implementation details were described in the original paper but are not repeated here verbatim by the source abstract.
## Clinical Analysis & Structured Key Points
[ Skip to main content ](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114#main-content) Advertisement * [plos.org](https://plos.org/) * [Create account](https://community.plos.org/registration/new) * [Sign in](https://journals.plos.org/user/secure/login?page=%2Fplosone%2Farticle%3Fid%3D10.1371%2Fjournal.pone.0361114) * * About * Browse * Publish * [](https://journals.plos.org/plosone/ "PLOS One") * Search [advanced search](https://journals.plos.org/plosone/search) * [Browse Topics](https://journals.plos.org/plosone/subjectAreaBrowse) Browse Subject Areas ? Click through the PLOS taxonomy to find articles in your field. For more information about PLOS Subject Areas, click [here](https://github.com/PLOS/plos-thesaurus/blob/master/README.md "Link opens in new window"). [](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114) [](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114) * 0 [Save](https://journals.plos.org/plosone/article/metrics?id=10.1371/journal.pone.0361114#savedHeader) [Total Mendeley and Citeulike bookmarks.](https://journals.plos.org/plosone/article/metrics?id=10.1371/journal.pone.0361114#savedHeader) * 0 [Citation](https://journals.plos.org/plosone/article/metrics?id=10.1371/journal.pone.0361114#citedHeader) [Paper's citation count computed by Dimensions.](https://journals.plos.org/plosone/article/metrics?id=10.1371/journal.pone.0361114#citedHeader) * 0 [View](https://journals.plos.org/plosone/article/metrics?id=10.1371/journal.pone.0361114#viewedHeader) [PLOS views and downloads.](https://journals.plos.org/plosone/article/metrics?id=10.1371/journal.pone.0361114#viewedHeader) * 0 [Share](https://journals.plos.org/plosone/article/metrics?id=10.1371/journal.pone.0361114#discussedHeader) [Sum of Facebook, Twitter, Reddit and Wikipedia activity.](https://journals.plos.org/plosone/article/metrics?id=10.1371/journal.pone.0361114#discussedHeader) Open Access Peer-reviewed Research Article # TCLNet: A Thermal Contrast-Aware Lightweight Network for infrared object detection * Xiyuan Zhu , Roles Conceptualization, Data curation, Formal analysis, Funding acquisition * E-mail: zhu19806198591@163.com Affiliation Xijing University, Xi’an, Shaanxi, China [ ![ORCID logo](https://journals.plos.org/resource/img/orcid_16x16.png) https://orcid.org/0009-0008-1723-9284 ](https://orcid.org/0009-0008-1723-9284 "ORCID Registry") ⨯ * Hongtao Dang, Roles Conceptualization Affiliation Xijing University, Xi’an, Shaanxi, China ⨯ * Xiaoyuan Jin Roles Conceptualization Affiliation Xijing University, Xi’an, Shaanxi, China ⨯ # TCLNet: A Thermal Contrast-Aware Lightweight Network for infrared object detection * Xiyuan Zhu, * Hongtao Dang, * Xiaoyuan Jin ![PLOS](https://journals.plos.org/resource/img/logo-plos-full-color.svg) x * Published: October 22, 2026 * * [Article](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114) * [Authors](https://journals.plos.org/plosone/article/authors?id=10.1371/journal.pone.0361114) * [Metrics](https://journals.plos.org/plosone/article/metrics?id=10.1371/journal.pone.0361114) * [Comments](https://journals.plos.org/plosone/article/comments?id=10.1371/journal.pone.0361114) * [Media Coverage](http://plos.altmetric.com/details/doi/10.1371/journal.pone.0361114) * [Abstract](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114#abstract0) * [1. Introduction](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114#sec001) * [2. Related work](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114#sec002) * [3. Method](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114#sec006) * [4. Research methods](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114#sec011) * [5. Experiments and analysis](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114#sec015) * [6. Conclusion](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114#sec021) * [References](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114#references) * [Reader Comments](https://journals.plos.org/plosone/article/comments?id=10.1371/journal.pone.0361114) * [Figures](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114) ## Abstract Infrared object detection plays an important role in low-light and complex-environment perception. However, infrared images often exhibit low signal-to-noise ratio, weak target-background contrast, and limited texture information, which can degrade detection performance in small-target and cluttered-background scenarios while making it difficult to balance detection accuracy and model complexity. To address these challenges, this paper proposes TCLNet, a Thermal Contrast-Aware Lightweight Network for infrared object detection. TCLNet introduces a Thermal Contrast Encoder (TCE) to enhance local contrast-sensitive target representations and a Thermal-Guided Cross-Feature Attention (TGCA) mechanism to promote interactions between contrast-enhanced and semantic features, thereby improving target discrimination and suppressing background interference.Experiments were conducted independently on the M3FD, FLIR ADAS v2 (denoted as FLIR_2), and NUDT-SIRST1 datasets. TCLNet achieves mAP50 values of 60.5%, 45.4%, and 91.4%, respectively, corresponding to improvements of 12.8, 11.3, and 7.5 percentage points over the YOLO12n baseline. Meanwhile, the model maintains a compact size of approximately 2M parameters. These results demonstrate that TCLNet improves infrared target detection, particularly in low-SNR and small-target conditions, while achieving a favorable balance between detection accuracy and model complexity. ## Figures ![Fig 6](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361114.g006) ![Table 17](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361114.t017) ![Fig 7](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361114.g007) ![Fig 1](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361114.g001) ![Fig 2](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361114.g002) ![Table 1](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361114.t001) ![Fig 3](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361114.g003) ![Table 2](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361114.t002) ![Table 3](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361114.t003) ![Table 10](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361114.t010) ![Table 4](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361114.t004) ![Table 5](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361114.t005) ![Table 6](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361114.t006) ![Table 7](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361114.t007) ![Table 8](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361114.t008) ![Table 9](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361114.t009) ![Table 11](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361114.t011) ![Table 12](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361114.t012) ![Table 13](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361114.t013) ![Table 14](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361114.t014) ![Table 15](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361114.t015) ![Table 16](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361114.t016) ![Fig 4](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361114.g004) ![Fig 5](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361114.g005) ![Fig 6](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361114.g006) ![Table 17](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361114.t017) ![Fig 7](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361114.g007) ![Fig 1](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361114.g001) ![Fig 2](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361114.g002) ![Table 1](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361114.t001) **Citation:** Zhu X, Dang H, Jin X (2026) TCLNet: A Thermal Contrast-Aware Lightweight Network for infrared object detection. PLoS One 21(10): e0361114. https://doi.org/10.1371/journal.pone.0361114 **Editor:** Yongjie Li, University of Electronic Science and Technology of China, CHINA **Received:** July 15, 2026; **Accepted:** October 6, 2026; **Published:** October 22, 2026 **Copyright:** © 2026 Zhu et al. This is an open access article distributed under the terms of the [Creative Commons Attribution License](http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. **Data Availability:** No new primary image dataset was generated in this study. All experiments were conducted using three publicly available third-party datasets: the Multi-Spectral Multi-Scenario Fusion Dataset (M3FD), the Teledyne FLIR ADAS Thermal Dataset v2, and the Single-frame InfraRed Small Target (SIRST) dataset. The M3FD dataset is publicly available through the official TarDAL repository at: The repository provides the dataset description and download links for the M3FD infrared-visible image pairs and corresponding annotations. The Teledyne FLIR ADAS Thermal Dataset v2 is available from the official Teledyne FLIR dataset portal at: Access to the FLIR dataset requires submission of the registration form and acceptance of the data provider’s applicable terms. Other researchers can obtain the dataset directly from Teledyne FLIR through the same procedure followed by the authors. The SIRST dataset, including the infrared images and corresponding annotations, is publicly available through the official repository at: . **Funding:** The author(s) received no specific funding for this work. **Competing interests:** The authors have declared that no competing interests exist. ## 1. Introduction Infrared object detection is important for nighttime monitoring, autonomous driving, industrial inspection, and intelligent transportation. Unlike visible-light cameras, infrared sensors record scene-dependent intensity patterns influenced by scene radiance, object properties, and sensor response. Infrared imaging can therefore provide complementary visual information under low illumination and certain adverse environmental conditions [[1](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114#pone.0361114.ref001),[2](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114#pone.0361114.ref002)]. Despite these advantages, infrared images present several challenges for object detection. First, targets often exhibit weak contrast relative to complex backgrounds because of low signal-to-noise ratio (SNR) or low signal-to-clutter ratio (SCR). Second, limited spatial resolution, sensor noise, and smooth intensity transitions can produce indistinct target boundaries. Third, the absence of chromatic information and the scarcity of detailed textures restrict the discriminative capacity of conventional feature representations. These limitations are particularly severe for small targets, which may occupy only a few pixels and can be progressively weakened during feature downsampling [[3](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114#pone.0361114.ref003)]. Infrared intensity patterns are related to scene radiance but should not be directly interpreted as calibrated temperature measurements. Their values are also influenced by sensor characteristics, object emissivity, environmental conditions, and image preprocessing [[4](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114#pone.0361114.ref004)]. Nevertheless, infrared targets frequently exhibit local intensity differences and spatial response variations relative to their surrounding backgrounds. These observable patterns motivate the use of relative-contrast and gradient-based feature cues. Such cues provide a thermally inspired inductive bias without requiring explicit recovery of temperature, radiance, or heat-diffusion parameters [[5](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114#pone.0361114.ref005),[6](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114#pone.0361114.ref006)]. Recent advances in deep learning have substantially improved general object detection. Existing approaches include two-stage convolutional detectors, such as the R-CNN family [[7](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114#pone.0361114.ref007)–[12](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114#pone.0361114.ref012)], one-stage detectors, such as the YOLO family [[13](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114#pone.0361114.ref013)–[17](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114#pone.0361114.ref017)], and transformer-based detectors represented by DETR and its variants [[18](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114#pone.0361114.ref018)–[21](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114#pone.0361114.ref021)]. Although these detectors perform well on visible-light benchmarks, their architectures are not specifically designed for the weak contrast and limited texture commonly observed in infrared images. Directly applying them to infrared data may produce suboptimal feature representations, particularly for small targets and cluttered backgrounds. To address these limitations, this paper proposes TCLNet, a Thermal Contrast-Aware Lightweight Network for infrared object detection. TCLNet is constructed using a lightweight YOLO-based framework and introduces contrast-sensitive feature enhancement and cross-feature attention. Instead of estimating physical thermal quantities, the proposed network learns relative target-background responses from infrared features. It then uses these responses to guide the interaction between contrast-enhanced and semantic representations. This design is intended to improve target discrimination while retaining a compact parameter count. The main contributions of this study are summarized as follows: 1. Thermal Contrast Encoder (TCE): A thermally inspired feature-enhancement module is introduced to characterize local target-background contrast and spatial response variations. TCE combines learned response projection, local background-response estimation, normalized relative-contrast calculation, and Sobel-gradient extraction. These operations are used to generate a contrast-sensitive modulation map for infrared feature enhancement. 2. Thermal-Guided Cross-Feature Attention (TGCA): A local cross-attention mechanism is developed to facilitate interaction between contrast-enhanced and semantic features. The contrast-enhanced features generate the Query, while the original semantic features generate the Key and Value. A local-window strategy reduces the computational cost of global attention and limits interactions with spatially distant background regions. 3. Thermal-Attentive Fusion C2f (TAF-C2f): TCE and TGCA are integrated into a lightweight feature-extraction block. The resulting module jointly optimizes contrast-sensitive enhancement, local cross-attention, and semantic feature extraction within a compact parameter budget. The remainder of this paper is organized as follows. Section 2 reviews related work on infrared object detection. Section 3 describes the TCLNet architecture and its main components. Section 4 presents the datasets, evaluation metrics, and implementation details. Section 5 reports the comparative experiments, ablation studies, visual analysis, and robustness evaluation. Section 6 concludes the paper and discusses possible directions for future research. ## 2. Related work ### 2.1. Traditional Infrared Target Detection Methods The early infrared target detection methods mainly relied on manual feature design and signal processing techniques, such as the detection method based on Haar features and AdaBoost cascade classifier [[22](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114#pone.0361114.ref022)], the HOG feature method [[23](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114#pone.0361114.ref023)], and the detection strategy based on local feature transformation [[24](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114#pone.0361114.ref024)]. These methods usually construct specific feature descriptors to distinguish targets from backgrounds, achieving certain results in simple scenarios. However, due to the common characteristics of infrared images, such as low contrast, weak texture, and low signal-to-noise ratio, traditional artificial features often have difficulty accurately depicting the complex feature distribution of infrared targets. Moreover, these methods are sensitive to environmental changes and complex backgrounds, and are prone to false detections or missed detections in strong noise or cluttered environments, thus having significant limitations in robustness and generalization in complex infrared scenarios. To improve the performance of infrared target detection, some studies have introduced statistical modeling and signal processing methods to separate targets from backgrounds. For example, Gao et al. [[25](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114#pone.0361114.ref025)] proposed an infrared patch image model, achieving target-background separation through low-rank sparse decomposition, but this method has high computational complexity and is difficult to meet real-time requirements in practical applications. Chen et al. [[26](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114#pone.0361114.ref026)] proposed a local contrast detection method based on the human visual system (HVS), which to some extent alleviated the missed detection problem of low-signal-to-noise ratio targets, but still had a relatively high false alarm rate in complex cluttered backgrounds. Guo et al. [[27](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361114#pone.0361114.ref027)] proposed a frequency-domain saliency detection method based on phase spectrum, improving detection efficiency through frequency-domain analysis, but single-scale analysis is prone to
## Related Clinical Research

- [Intratracheal microspray of iohexol prolongs lung enhancement for preclinical micro-CT in mice](https://medichelpline.com/clinical-feed/plos-one-18-intratracheal-microspray-delivery-of-iohexol-for-sustained-lung-parenchymal.md)
- [NICE recommends five AI tools to support prostate cancer detection on MRI](https://medichelpline.com/clinical-feed/bmj-0-five-ai-tools-for-prostate-cancer-screening-are-recommended-by-nice.md)
- [Multimodal AI Detects In-Hospital Lung Cancer and Heart Failure from Routine Data](https://medichelpline.com/clinical-feed/medrxiv-11-explainable-multimodal-ai-for-in-hospital-detection-of-lung-cancer-and-heart.md)
- [Ordinal Multi-Instance Learning Automates 25-Task Lumbar Degeneration Grading on Sagittal MRI](https://medichelpline.com/clinical-feed/medrxiv-13-ordinal-multi-instance-learning-for-automated-multi-task-grading-of-lumbar.md)
- [X-ray diffraction of hydrated macromolecular crystals using a laser-driven X-ray source](https://medichelpline.com/clinical-feed/biorxiv-23-x-ray-diffraction-from-hydrated-macromolecular-crystals-using-laser-driven-x.md)

## Navigation
- [← Back to Radiology Feed](https://medichelpline.com/clinical-feed/radiology.md)
- [← All Clinical Specialties](https://medichelpline.com/clinical-feed.md)
## Medical & Regulatory Disclaimer

> [!CAUTION]
> MedicHelpline content is structured for research, educational, and professional discovery purposes. It does not constitute individual medical advice, clinical diagnosis, or treatment recommendations.
> Always verify dosing, contraindications, and regulatory alerts against official product labeling and primary regulatory sources before clinical decision-making.