Low Illumination Image Object Detection Method Based on ICFIE-YOLO

Images obtained in low light environments often have low brightness, low contrast, and uneven lighting, re⁃ sulting in weakened and blurred image features that are difficult to extract. At the same time, there is also a large amount of noise information in the limited extracted features, making it difficult to detect and recognize objects. Therefore, there are very few existing low light object detection results. This paper proposes a low illumination object detection method based on the Illumination Correction and Feature Interaction Enhancement (ICFIE-YOLO) network to address the difficulties in extracting features from low illumination objects and the large noise in the feature space. This method first utilizes the pro⁃ posed ICFIE-YOLO internal Multi Scale Illumination Correction Network (MSICN) to correct low illumination images, highlighting the blurry features of objects hidden in the image’s background, and enabling the feature extraction module to better extract object features. Secondly, to fully utilize effective feature information and filter out noise interference in fea⁃ ture maps, a Feature Interacted Enhancement (FIE) detection head is proposed. Through feature attention interaction, feature enhancement is achieved, establishing spatial and semantic correlations between features in different regions of low illumi⁃ nation images, thereby suppressing the interference of noise on effective features and achieving feature enhancement. Final⁃ ly, on the basis of enhancing features and removing noise, an improved detection head is used to achieve high-precision ob⁃ ject detection. Experiments on the ExDark and DarkFace datasets show that the proposed Method improves mAP by over 2.1 percentage points compared to mainstream object detection models, increases recall by over 4.2 percentage points com⁃ pared to existing low light object detection Methods, and improves recall by 2.6 percentage points compared to baseline models. The proposed Method has good generalization performance. © 2025 Chinese Institute of Electronics. All rights reserved.