💻 computer science

AMIEOD: Adaptive Multi-Experts Image Enhancement for Object Detection in Low-Illumination Scenes

The paper proposes AMIEOD, a joint image enhancement and object detection framework for low-illumination scenes that utilizes a Multi-Experts Image Enhancement Module guided by detection-oriented losses to dynamically select optimal enhancement strategies and significantly improve detection accuracy.

Original authors: Xiaochen Huang, Honggang Chen, Weicheng Zhang, Xiaobo Dai, Yongyi Li, Linbo Qing, Xiaohai He

Published 2026-05-08

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Xiaochen Huang, Honggang Chen, Weicheng Zhang, Xiaobo Dai, Yongyi Li, Linbo Qing, Xiaohai He

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 friend in a crowded, pitch-black room. If you just squint your eyes (the standard computer vision approach), you might miss them or mistake a coat rack for a person. This is the problem computers face when trying to "see" objects in low-light conditions, like at night or in dark corners.

The paper introduces a new system called AMIEOD (Adaptive Multi-Experts Image Enhancement for Object Detection). Think of it not as a single pair of glasses, but as a smart team of specialists who work together to help a security guard (the object detector) find people and things in the dark.

Here is how it works, broken down into simple concepts:

1. The Problem: One Size Doesn't Fit All

Usually, when we try to fix a dark photo, we use a single tool (like a "brightness" slider). But the paper argues that different dark scenes need different fixes. A dark room with a flashlight needs a different fix than a foggy night street. Using just one tool often leaves some details blurry or creates weird artifacts (glitches) that confuse the computer.

2. The Solution: A "Team of Experts" (MEIEM)

Instead of relying on one tool, AMIEOD brings in a Multi-Experts Image Enhancement Module (MEIEM). Imagine a workshop with three different specialists:

Expert A (The Pre-trained Pro): Uses a tried-and-true method to brighten things up quickly.
Expert B (The Joint Learner): A specialist who learns specifically how to brighten the image while watching the security guard to make sure the brightening helps the guard find the target.
Expert C (The Adaptive Artist): A specialist who looks at the specific picture and says, "This one needs more contrast," or "This one needs better color balance," and adjusts the settings on the fly.

During training, the system feeds the dark image to all three experts at once. The security guard tries to find the objects in all three resulting versions.

3. The Coach: "Detection-Guided Regression Loss" (DGRL)

How does the team know which expert is doing the best job?
The paper introduces a "Coach" (DGRL). The Coach looks at the results of all three experts. If Expert B's version helps the guard find the most objects, the Coach says, "Okay, everyone else, try to make your output look more like Expert B's output."

The Metaphor: It's like a choir where the conductor listens to the best singer and tells the others to match their pitch. This ensures the whole team gets better at the specific goal: helping the guard detect objects, not just making the picture look pretty.

4. The Smart Switch: "Expert Selection Module" (ESM)

Here is the tricky part: When the system is actually working in the real world (inference), it can't use all three experts at once because that would be too slow. It has to pick one.

The Metaphor: Imagine a bouncer at a club who has to decide which VIP guest gets in. The bouncer (ESM) looks at the incoming dark image and asks, "Is this a foggy night? A backlit room? A shadowy alley?" Based on the "vibe" of the image, the bouncer instantly picks the single best expert from the team to handle that specific job.
This decision is made using a special scoring system (DGCE loss) that was trained to know which expert works best for which type of darkness.

5. The Results: Better Vision in the Dark

The authors tested this system on several datasets (collections of dark photos) and compared it to other methods.

The Outcome: Their "Team of Experts" approach significantly outperformed the old methods. It found more objects (higher accuracy) and missed fewer targets than previous systems.
Speed: Even with the extra "bouncer" and "team" logic, the system is still fast enough to work in real-time (about 70 frames per second), which is crucial for things like autonomous driving or security cameras.

Summary

In short, AMIEOD is a smart system that doesn't just blindly brighten a dark image. Instead, it:

Consults a team of different enhancement methods.
Trains them together so they all focus on helping the object detector succeed.
Picks the perfect specialist for each specific image before making the final decision.

This approach allows computers to "see" much better in the dark, making them more reliable for safety and surveillance tasks.

Technical Summary: AMIEOD – Adaptive Multi-Experts Image Enhancement for Object Detection in Low-Illumination Scenes

Problem Statement
Object detection systems trained on standard, well-lit datasets often suffer significant performance degradation when applied to low-illumination scenarios (e.g., night, backlighting, short exposure). Existing solutions generally fall into three categories: Low-Light Image Enhancement (LLIE) followed by detection, Jointly Optimized Enhancement and Detection (JED), and Illumination-Invariant Feature Extraction (IIFE). However, these approaches often rely on single improvement schemes that fail to adapt flexibly to the diverse characteristics of images across different scenes and imaging conditions. Furthermore, a domain gap often persists between enhanced images and the data distribution expected by detection models, limiting the effectiveness of visual perception in dark environments.

Methodology
The authors propose AMIEOD (Adaptive Multi-Experts Image Enhancement for Object Detection), a framework that jointly optimizes image enhancement and object detection in a detection-performance-oriented manner. The framework consists of three core components:

Multi-Experts Image Enhancement Module (MEIEM):
Instead of relying on a single enhancement strategy, MEIEM integrates three distinct sub-modules to extract complementary information from low-light images:
- PIEM (Pretrained Image Enhancement Module): Utilizes a frozen SCI (Self-Calibrated Illumination) model to enhance brightness and provide clearer object cues.
- JIEM (Jointly Optimized Enhancement Module): Shares the architecture of PIEM but is jointly optimized with the detector to ensure the enhancement aligns with detection-aware representations.
- IAEM (Image-Adaptive Enhancement Module): Based on IAYOLO, this module uses a Parameter Prediction (PP) network to predict adaptive parameters for a Differentiable Image Processing (DIP) module, adjusting contrast, white balance, and tone based on global image characteristics.
  During training, the input image is processed by all experts to generate multiple enhanced variants ( $I_1, I_2, ..., I_n$ ) alongside the original, which are fed into the detector.
Detection-Guided Regression Loss (DGRL):
To ensure the sub-modules learn information beneficial for detection, DGRL is introduced. It identifies the enhanced output with the lowest detection loss among the original and all enhanced variants as the regression target. Other experts are then regressed toward this optimal target. This process minimizes the variance of multi-expert outputs around the optimal solution and suppresses noise during early training, effectively guiding the enhancement modules without interfering with the detector's optimization.
Expert Selection Module (ESM) with Detection-Guided Cross-Entropy (DGCE):
Since only one enhancement strategy can be applied during inference, an ESM is constructed to adaptively select the most suitable expert for a given input image. The ESM is trained using DGCE, which treats expert selection as a classification task. The ground-truth label for training is the expert that yields the minimum detection loss for a specific image. This allows the ESM to learn a policy that approximates the posterior probability of the optimal expert, dynamically routing the input to the best preprocessing strategy.

Key Contributions

Multi-Expert Framework: The introduction of MEIEM, which leverages diverse enhancement strategies (pretrained, jointly optimized, and image-adaptive) to comprehensively explore low-illumination image characteristics.
Detection-Guided Optimization: The proposal of DGRL, which optimizes the enhancement modules based on detection results, ensuring the enhanced images are tailored to the detection task.
Adaptive Routing: The development of the ESM and DGCE loss, enabling the system to dynamically select the most appropriate enhancement strategy for each input image based on its global characteristics, rather than using a static or single-scheme approach.
Generalizability: The method is designed to be detector-flexible, demonstrated to work effectively across various YOLO architectures (v3, v5, v7, v9).

Experimental Results
Extensive experiments were conducted on the ExDark dataset, with additional validation on DarkFace, LLVIP, and M3FD.

Performance: On the ExDark dataset, AMIEOD achieved an mAP of 82.1%, outperforming the baseline YOLOv3 by 5.6% and surpassing the current state-of-the-art method (EMV-YOLO) by 2.4%. It achieved the highest Average Precision (AP) in 9 out of 12 object categories.
Generalization: The method consistently improved performance across different datasets and detector backbones (YOLOv3, v5s, v7, v9), demonstrating robustness in diverse low-illumination conditions.
Efficiency: Despite the additional complexity of the ESM, the framework maintains real-time inference speeds of approximately 70 FPS at 640×640 resolution, offering a favorable trade-off between accuracy and efficiency.
Ablation Studies: Results confirmed that the combination of multi-expert training, DGRL, and adaptive expert selection (DGCE) is essential for the observed performance gains. Specifically, DGRL significantly accelerated convergence and stabilized training, while DGCE enabled the model to outperform both fixed and random expert selection strategies.

Significance and Claims
The paper claims that AMIEOD addresses the limitations of single-scheme enhancement methods by introducing a flexible, adaptive framework that aligns image enhancement directly with detection objectives. By jointly optimizing diverse enhancement experts and employing a detection-guided selection mechanism, the method significantly improves object detection accuracy in dim scenes. The authors emphasize that their approach effectively preserves structural and semantic information, enhancing robustness for small, dimly lit, or partially occluded objects. While acknowledging challenges with visually ambiguous or camouflaged objects, the work establishes a principled alternative to heuristic routing, offering a theoretically grounded approach to minimizing expected detection risk in low-illumination environments.