💻 computer science

IR-PowerDet: An Infrared Power Equipment Detection Algorithm Based on Multi-Scale Feature Fusion and Dynamic Environment Adaptation

This paper proposes IR-PowerDet, a YOLOv12s-based framework incorporating multi-scale feature fusion, dynamic environment adaptation, and small-object enhancement modules to achieve robust, high-accuracy detection of infrared power equipment in complex and occluded scenarios, reaching a 96.8% mAP.

Original authors: Yongbo Tang, Huicong Sun, Zijun Zhong, Miao Li

Published 2026-08-10

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Yongbo Tang, Huicong Sun, Zijun Zhong, Miao Li

Original paper licensed under CC BY 4.0 (https://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 find a tiny, specific toy hidden inside a giant, messy sandbox. Now, imagine that the sandbox is glowing with heat instead of being made of sand, and the "toy" is actually a piece of electrical equipment that might be getting too hot and causing a fire. This is the daily challenge for engineers who monitor power grids. They use special cameras that see heat (infrared) instead of light to spot problems before they happen. But these cameras have a hard time: the background is often cluttered, the "toys" (equipment) can be very small, and sometimes they are hidden behind other things. It's like trying to find a specific warm cookie in a pile of warm rocks. For a long time, computers struggled to do this quickly and accurately, often getting confused by the heat of the rocks or missing the tiny cookie entirely.

This is where a new team of researchers steps in with a clever solution. They built a smart computer brain, based on a famous type of artificial intelligence called YOLO (which stands for "You Only Look Once"), but they gave it a major upgrade specifically for these hot, messy power grid scenes. Think of their new system, named IR-PowerDet, as a super-powered detective that doesn't just look at the picture; it learns to ignore the distracting heat of the background, zooms in on tiny details it usually misses, and adapts its "eyes" depending on whether the scene is bright, dark, or foggy. They tested this detective on a custom collection of 1,200 real-world heat images of power equipment. The result? Their new detective found the equipment with incredible accuracy, catching 96.8% of the targets correctly, while running fast enough to watch the grid in real-time without getting tired.

The Problem: The Heat of the Moment

Power lines and transformers are the veins of our modern world, and keeping them safe is a big deal. If a piece of equipment gets too hot, it can fail, causing blackouts or even fires. Traditionally, humans had to walk around and check these things, or use old-school computer tricks that weren't very good at handling complex scenes. In recent years, we've started using deep learning—computer brains that learn by looking at thousands of pictures—to do the spotting. However, infrared images are tricky. They are full of "noise" (random heat from the sky or ground), and the equipment we need to find is often tiny or partially hidden. Standard computer brains often miss these small targets or get confused by the background, leading to false alarms or missed dangers.

The Solution: IR-PowerDet's Three Superpowers

The researchers at Chongqing University of Posts and Telecommunications didn't just take an existing AI model and hope for the best. They took a solid, fast model called YOLOv12s and added three special "modules" (think of them as new tools in a detective's kit) to make it perfect for infrared power inspection.

1. The Multi-Scale Detective (MSED)
Imagine trying to spot a tiny ant and a giant elephant in the same photo. A normal camera might focus on one and miss the other. The Multi-Scale Equipment Detection (MSED) module is like a detective who can look at the whole picture and zoom in on tiny details at the same time. It uses three different "lenses" (convolution branches) to look at the image: one that sees the big picture, one that focuses on fine details, and one that keeps the original view. By combining these views, the system can spot both large transformers and tiny, distant components without getting confused by the clutter.

2. The Environment Chameleon (DEAM)
Infrared images change a lot depending on the weather, the time of day, or if the sun is hitting a rock nearby. A rigid system might get confused when the background gets brighter or darker. The Dynamic Environment Adaptation Mechanism (DEAM) is like a chameleon that adjusts its eyes to the current lighting. It has two parts: one that looks at where things are in the picture (spatial attention) and one that looks at what kind of heat signals are important (channel attention). It learns to ignore the "noise" of the background heat and only pay attention to the specific heat signatures of the power equipment, making it much harder to trick.

3. The Small-Object Magnifier (SGPED)
This is the most critical tool for finding the "tiny cookies." When a computer looks at an image, it often shrinks it down to make it easier to process, which can make small objects disappear. The Small-object Guided Power Equipment Detection Enhancer (SGPED) is like a magic magnifying glass. It takes the deep, processed information and brings it back together with the original, high-detail "shallow" layers of the image. This helps the system remember the edges and shapes of small, blurry, or partially hidden equipment that would otherwise vanish into the background.

The Results: Speed and Smarts

The team tested their new detective, IR-PowerDet, on a dataset of 1,200 infrared images they collected themselves. These images were tricky, featuring different lighting, cluttered backgrounds, and equipment that was small or blocked.

The results were impressive. The new system achieved a 96.8% accuracy rate (measured as mAP), which means it correctly identified the equipment almost every time. It also ran incredibly fast, processing 93 frames per second (FPS). To put that in perspective, it was much faster than other popular models like Faster R-CNN (which ran at 11 FPS) or even other YOLO versions. It also stayed lightweight, using only 11.9 million parameters (the "brain size" of the model), which is smaller than many other high-accuracy models.

In their tests, the researchers showed that removing any of the three special tools (MSED, DEAM, or SGPED) made the system worse. For example, without the "Small-Object Magnifier," the system missed more tiny targets. Without the "Environment Chameleon," it got confused by background heat. But when all three worked together, the system became a powerhouse, outperforming existing methods in both speed and accuracy.

Why This Matters

This isn't just about making a cooler computer program; it's about keeping the lights on and preventing disasters. By making infrared detection faster and more accurate, this technology could be installed on drones or edge devices (small computers) that patrol power grids in real-time. Instead of waiting for a human to spot a hot wire, the system could instantly flag the problem, even if the equipment is small, hidden, or the weather is bad. The researchers suggest that while this is a big step forward, there is still work to be done to make it even better at handling new, unseen situations, but for now, IR-PowerDet offers a very promising way to keep our power systems safe and smart.

Technical Summary: IR-PowerDet

Problem Statement

The rapid advancement of intelligent power systems necessitates real-time infrared monitoring to ensure the safety and stability of power infrastructure. However, detecting power equipment in infrared imagery presents significant challenges that traditional methods and existing deep learning models struggle to address simultaneously. Key difficulties include:

Complex Backgrounds: Infrared images often contain substantial noise and cluttered environments.
Small Object Detection: Power equipment targets are frequently small, occupying minimal space in deep feature maps, leading to weakened semantic representation.
Occlusion and Blur: Targets are often partially occluded or have blurred boundaries, making localization difficult.
Dynamic Conditions: Variations in illumination and thermal backgrounds cause image quality fluctuations, resulting in false positives and missed detections.
Efficiency vs. Accuracy: While two-stage detectors (e.g., Faster R-CNN) offer accuracy, they lack the inference speed required for real-time monitoring. Conversely, one-stage detectors often sacrifice accuracy in complex, small-target scenarios.

Methodology

The paper proposes IR-PowerDet, a detection framework built upon the YOLOv12s baseline. The architecture is designed to balance detection accuracy, model complexity, and inference speed through three specific lightweight modules integrated into the backbone and neck of the network:

1. Multi-Scale Equipment Detection (MSED) Module

Designed to enhance feature fusion across different scales, the MSED module addresses feature degradation caused by background noise and occlusion. It employs a three-branch parallel architecture:

Multi-scale Dilated Convolution Branch: Uses 3×3 kernels with dilation rates of $r=1, 3, 5$ to capture varying receptive fields (3×3, 7×7, 11×11) without increasing parameters or losing spatial resolution.
Local Feature Compression Branch: Utilizes 1×1 convolutions to compress channel dimensions, reducing redundancy and enhancing the semantic representation of small targets.
Identity Mapping Branch: Preserves original input features to maintain information integrity and facilitate gradient propagation.
The outputs of these branches are concatenated to form a unified multi-scale feature representation.

2. Dynamic Environment Adaptation Mechanism (DEAM)

To suppress background interference and adapt to varying thermal conditions, the DEAM module replaces fixed attention strategies with a dynamic, task-oriented approach. It consists of two branches:

Infrared Spatial Dependency Branch: Models long-range dependencies between equipment regions and surrounding contexts using a spatial attention map derived from matrix multiplication and softmax operations. This enhances visibility under uneven illumination.
Thermal-Channel Response Recalibration Branch: Emphasizes feature channels sensitive to thermal responses while suppressing redundant background channels via self-correlation.
The fusion of these branches allows the model to adaptively recalibrate feature responses based on the specific infrared scene.

3. Small-object Guided Power Equipment Detection Enhancer (SGPED)

This module specifically targets the loss of detail in small targets caused by downsampling. It operates via cross-layer feature compensation:

Feature Recovery: An intermediate feature map (80×80) is compressed and upsampled to 160×160.
Cross-layer Fusion: The upsampled features are added to shallow feature maps of the same resolution, combining fine-grained edge details with high-level semantic information.
Refinement: The fused map undergoes convolutional operations to generate high-resolution prediction features, improving the localization of small, blurred, or partially occluded targets.

Key Contributions

The authors summarize their contributions as follows:

Framework Proposal: The introduction of IR-PowerDet, an efficient YOLOv12s-based framework tailored for complex infrared power inspection conditions.
Novel Modules: The design and integration of three lightweight modules (MSED, DEAM, and SGPED) to specifically improve multi-scale feature representation, environmental adaptability, and small-object detection capabilities.
Performance Validation: Demonstration of superior performance over the baseline and other mainstream detectors in terms of accuracy, parameter count, and inference speed.

Experimental Results

Experiments were conducted on a self-built dataset of 1,200 high-resolution infrared images (640×640) covering diverse operating conditions. The model was trained for 200 epochs using an NVIDIA RTX 4090.

Accuracy: IR-PowerDet achieved a Precision of 96.9%, Recall of 95.6%, and an mAP@0.5 of 96.8%.
Efficiency: The model operates at 93 FPS with only 11.9M parameters.
Comparative Performance:
- Outperformed SKS-YOLO (89.4% mAP, 55 FPS).
- Outperformed YOLOv5s-CASP (85.6% mAP, 66 FPS).
- Outperformed Faster R-CNN (80.1% mAP, 11 FPS).
- Outperformed SSD (73.3% mAP, 30 FPS).
Ablation Studies: Sequential addition of MSED, DEAM, and SGPED to the YOLOv12s baseline resulted in consistent performance gains, with the full model achieving a 14.9% improvement in mAP over the baseline (81.9% → 96.8%).

Significance and Claims

The paper claims that IR-PowerDet provides a practical solution for intelligent power inspection by effectively addressing the trade-off between detection accuracy and real-time inference speed. Its significance lies in:

Robustness: The ability to maintain high accuracy in complex infrared scenarios involving small targets, occlusion, and dynamic thermal backgrounds.
Deployability: The lightweight nature of the model (11.9M parameters) and high inference speed (93 FPS) make it suitable for deployment on edge devices within smart grids and safety monitoring systems.
Advancement: The work moves beyond standard YOLO implementations by introducing specific mechanisms for infrared-specific challenges, such as thermal background suppression and small-target feature recovery.

The authors conclude that while the current results are promising, future work will explore few-shot learning, cross-domain adaptation, and model compression to further enhance generalization in even more complex conditions.