💻 computer science

CarboFormer: A Lightweight Semantic Segmentation Architecture for Efficient Carbon Dioxide Detection Using Optical Gas Imaging

This paper introduces CarboFormer, a lightweight semantic segmentation architecture that achieves high-accuracy, real-time carbon dioxide detection in optical gas imaging for diverse applications like industrial leak monitoring and livestock management, while operating efficiently on resource-constrained platforms.

Original authors: Taminul Islam, Toqi Tahamid Sarker, Mohamed G Embaby, Khaled R Ahmed, Amer AbuGhazaleh

Published 2026-05-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Taminul Islam, Toqi Tahamid Sarker, Mohamed G Embaby, Khaled R Ahmed, Amer AbuGhazaleh

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to spot a faint, invisible cloud of carbon dioxide (CO2) drifting through the air. To the naked eye, it's invisible. Even to a standard camera, it's just empty space. But to a special "heat-sensing" camera (called Optical Gas Imaging or OGI), this gas looks like a ghostly, swirling plume of smoke.

The problem is that these gas clouds are tricky. Sometimes they are thick and obvious; other times, they are so thin and slow-moving (like a slow leak from a cow's stomach or a tiny pipe) that they are incredibly hard to see.

Enter CarboFormer.

Think of CarboFormer as a super-smart, lightweight pair of glasses designed specifically for a robot drone. Its job is to look at the video feed from that special heat camera and instantly draw a perfect outline around the CO2 cloud, separating the "gas" from the "background."

Here is how the paper explains this technology in simple terms:

1. The Two "Training Camps" (The Datasets)

To teach this AI how to see gas, the researchers built two specific "training camps" (datasets):

The "Controlled Leak" Camp (CCR): Imagine a lab where they release pure CO2 from a tank at different speeds, from a tiny trickle (10 units) to a strong stream (100 units). They took thousands of photos of these leaks to teach the AI what different flow rates look like.
The "Cow Stomach" Camp (RTA): This is the real-world test. They used a machine that simulates the fermentation inside a dairy cow's stomach. As the cow's rumen fluid breaks down food, it releases CO2. They filmed these real biological emissions to see if the AI could spot gas coming from a living animal, not just a tank.

2. The "Lightweight" Brain

Most AI models that do this kind of work are like heavy, powerful supercomputers. They are accurate, but they are too big and slow to fit on a small drone that needs to fly around a farm or factory.

CarboFormer is different. The authors designed it to be "lightweight."

The Metaphor: Imagine a marathon runner. A heavy supercomputer is like a runner carrying a backpack full of bricks. It might be strong, but it's slow. CarboFormer is like a runner who has shed all the unnecessary weight but kept the muscle. It runs just as fast (or faster) and is much more agile, making it perfect for a drone that needs to move quickly.

3. How It Works (The Secret Sauce)

The paper highlights three main tricks the AI uses to see the gas so well:

Adaptive Vision: It knows how to look at the gas differently depending on how fast it's moving. If the gas is a slow, thin wisp, the AI zooms in on tiny details. If it's a thick, fast cloud, it steps back to see the whole picture. It's like a photographer who automatically switches lenses depending on the subject.
The "Helper" Brain (Auxiliary Supervision): During training, the AI has a "helper" branch that checks its work at an intermediate stage. It's like a teacher walking around a classroom, giving quick hints to students while they are still working, rather than just grading them at the end. This helps the AI learn faster and more accurately, especially for those tricky, slow-moving leaks.
Multi-Scale Mixing: It combines different views of the image (close-up and far-away) to make sure it doesn't miss the edges of the gas cloud.

4. The Results: Fast and Accurate

The researchers tested CarboFormer against other famous AI models (like SegFormer and SegNeXt).

The Score: On the "Controlled Leak" test, CarboFormer got an 84.88% accuracy score. On the "Cow Stomach" test, it scored an impressive 92.98%.
The Speed: It processes video at 84.68 frames per second. That is fast enough to watch a live video without any lag.
The Comparison: It beat other "lightweight" models and even competed with much larger, heavier models, but it did so using far less computer power.

5. Why This Matters

The paper concludes that this tool is a game-changer for precision livestock management (checking cow health) and industrial leak detection. Because it is so efficient, it can run on programmable drones.

In a nutshell: The researchers built a smart, fast, and small AI that can "see" invisible carbon dioxide gas in real-time. It's like giving a drone X-ray vision specifically for gas leaks, allowing it to fly around farms and factories to find emissions that humans and standard cameras would miss.

Technical Summary: CarboFormer

Problem Statement

Carbon dioxide (CO2) emissions are critical indicators of environmental impact and industrial processes, including livestock management. While traditional quantification methods like Fourier-transform infrared (FTIR) spectrometers exist, they are often expensive, cumbersome, and laboratory-centric, limiting their applicability in diverse field environments. Although recent advances in machine learning and computer vision have addressed CO2 capture optimization and gas detection, existing methods face limitations regarding specific datasets, high computational requirements, and applicability to real-world scenarios, particularly at low flow rates.

A significant research gap exists in the quantification of CO2 emissions using semantic segmentation combined with Optical Gas Imaging (OGI), specifically for agricultural sources like livestock. Current methods struggle to detect and accurately quantify the subtle thermal signatures of CO2 at minimal flow rates (e.g., 10–30 SCCM) and in complex thermal patterns at higher rates. Furthermore, no published research has utilized semantic segmentation to detect CO2 emissions from dairy cow rumen fluid, a critical area for precision livestock management.

Methodology

The authors propose CarboFormer, a lightweight semantic segmentation architecture designed to detect and quantify CO2 emissions from OGI imagery. The model is optimized for resource-constrained platforms, such as programmable drones, while maintaining competitive accuracy.

Architecture

CarboFormer integrates an optimized encoder-decoder architecture with three key innovations:

Adaptive Hierarchical Feature Extraction: The encoder employs a four-stage hierarchy with dimensions [32, 64, 160, 256]. It utilizes spatial reduction ratios [8, 4, 2, 1] and an overlap patch merging strategy to preserve spatial continuity. This design captures fine-grained structures in low-flow scenarios (where thermal contrast is <5%) and complex morphological patterns in high-flow conditions.
Multi-Scale Feature Integration: The decoder processes features through adaptive learnable combinations to integrate local boundary details with global plume structures. This ensures robust segmentation across diverse thermal intensities.
Robust Auxiliary Supervision: An auxiliary supervision branch is introduced at the third stage (S3, 160-dimensional feature level). This intermediate supervision captures discriminative patterns crucial for both subtle and pronounced emissions, enhancing learning consistency across varying flow intensities.

Training Strategy

The model utilizes a weighted dual-supervision loss function: $L = L_{main} + 0.4L_{aux}$ . Training employs the AdamW optimizer with a two-phase learning rate schedule (LinearLR warm-up followed by polynomial decay). Data augmentation includes random resizing and cropping to 512×512 pixels.

Datasets

To facilitate accurate CO2 plume segmentation, the authors curated two novel datasets:

Controlled Carbon Dioxide Release (CCR) Dataset: Captured using a FLIR G343 OGI camera under controlled conditions. It features systematic CO2 releases from a high-purity cylinder at flow rates ranging from 10 to 100 SCCM in 10 SCCM increments. The dataset contains 19,731 images (640×480 pixels) across multiple visualization modes.
Real Time Ankom (RTA) Dataset: Captures real-time CO2 emissions from dairy cow rumen fluid using the ANKOM gas production system. It includes 613 images (640×480 pixels) representing single-shot bursts under varying pH conditions. This dataset addresses the previously unexplored domain of livestock emissions.

A semi-automated pipeline using differential background modeling and the watershed algorithm was developed to generate pixel-precise segmentation masks for both datasets.

Key Contributions

CarboFormer Model: A lightweight semantic segmentation model explicitly designed for OGI imagery, optimized for resource-constrained platforms while maintaining high accuracy.
Novel Datasets: The introduction of the CCR and RTA datasets, with a specific focus on livestock emissions and low-flow scenarios.
Rigorous Evaluation: Comprehensive testing demonstrating significant improvements in detection accuracy and practical applicability, particularly in challenging low-flow conditions.

Experimental Results

The model was evaluated against state-of-the-art segmentation techniques (including SegFormer-B0, SegNeXt, DeepLabv3+, and Swin-T) on both datasets.

CCR Dataset Performance: CarboFormer achieved 84.88% mIoU, outperforming SegFormer-B0 (83.36%) and SegNeXt (82.55%). It maintained a computational footprint of 5.07M parameters and operated at 84.68 FPS, with a computational cost of 11.39 GFLOPs.
RTA Dataset Performance: Through transfer learning (pre-training on CCR and fine-tuning on RTA), CarboFormer achieved 92.98% mIoU and 96.07% mF-score, surpassing all compared models, including heavier architectures like DeepLabv3+ and Swin-T.
Qualitative Analysis: Visual comparisons demonstrated CarboFormer's superior boundary preservation and ability to detect complex plume morphologies across varying flow rates (10–100 SCCM), where baseline models often exhibited fragmentation or over-segmentation.
Ablation Studies: Experiments confirmed that the 3-layer transformer depth, 4-stage decoder, and auxiliary supervision head are essential for balancing accuracy and efficiency. Removing the auxiliary head reduced mIoU to 82.45%, while reducing decoder stages dropped mIoU to 80.13%.

Significance and Claims

The paper claims that CarboFormer advances environmental sensing and precision livestock management by providing robust, efficient tools for CO2 emission analysis. The model is particularly significant for:

Resource-Constrained Deployment: Its efficiency makes it suitable for real-time monitoring on programmable drones and mobile platforms.
Low-Flow Detection: It effectively addresses the challenge of detecting subtle thermal signatures in low-flow scenarios (10–30 SCCM), which is critical for agricultural applications.
Agricultural Innovation: By enabling the detection of CO2 from dairy cow rumen fluid, the work opens new possibilities for animal health monitoring through respiratory pattern analysis.

The authors acknowledge limitations, including the relatively small size of the real-world RTA dataset and potential for further accuracy improvements at very low flow rates. Future work is proposed to expand the RTA dataset with direct farm measurements, implement temporal modeling for dynamic gas patterns, and further enhance the efficiency-accuracy trade-off.