MAE-YOLO improves small object detection for intelligent inspection

Engineering notes

Firstly, a multi-scale edge space feature extraction module is proposed to optimize the edge and space feature extraction, which significantly improves the detection accuracy of small targets. By synergistically integrating MSESTE for edge-aware feature extraction, AMCFN for adaptive multi-scale context fusion, and ELCIN for parameter-efficient detection, MAE-YOLO achieves both lightweight design and high accuracy for small objects.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Abstract Intelligent inspection technology has become increasingly popular in industrial fields such as power facility maintenance (e.g., identifying cracked insulators, corroded transformers), traffic management (e.g., detecting vehicle anomalies), and industrial equipment upkeep (e.g., spotting surface defects on machinery, verifying the presence of small components like bolts and valves). The targets in these scenarios are often small-sized, well-defined objects that are critical to operational safety and efficiency. Although the traditional deep learning methods such as YOLO series have made progress in this field, however, in the detection of small targets under complex background, it still suffers from high false detection and miss rates, as well as high computational complexity, making it difficult to meet the real-time requirements of practical applications. In this article, an intelligent inspection model named MAE-YOLO is proposed. Firstly, a multi-scale edge space feature extraction module is proposed to optimize the edge and space feature extraction, which significantly improves the detection accuracy of small targets. Meanwhile, the adaptive multi-scale context fusion network is introduced to integrate the features of different scales effectively, thereby enhancing the robustness and adaptability of the model in dynamic environments. Finally, we propose an adaptive cavity shared detection head to further reduce the false detection and missing detection rate in multi-scale detection. By synergistically integrating MSESTE for edge-aware feature extraction, AMCFN for adaptive multi-scale context fusion, and ELCIN for parameter-efficient detection, MAE-YOLO achieves both lightweight design and high accuracy for small objects. Experimental results on the VisDrone2019 dataset show that the accuracy of MAE-YOLO is improved by 2.6% compared to the original YOLOv8n model. According to the results of the self-collected dataset, it can be seen that MAE-YOLO only needs 4.7 MB, which is reduced by 24% compared with YOLOv8n, while maintaining high detection accuracy. Unlike existing lightweight methods that often sacrifice edge details for efficiency, MAE-YOLO preserves fine-grained object boundaries through Sobel-based edge enhancement while reducing detection head parameters by 53%, achieving a superior accuracy-size trade-off (35.1% mAP@50 at 4.57 MB) compared to recent detectors such as SOD-YOLO (30.08% mAP@50). To facilitate reproducibility and further research, the source code of MAE-YOLO has been released at: https://github.com/970334745/MAE-YOLO .

Links and sources

• PDF from original source

Comments