ECAN-Detector: An Efficient Context-Aggregation Network for Small-Object Detection
Over the past decade, the field of object detection has advanced remarkably, especially in the accurate recognition of medium- and large-sized objects. Nevertheless, detecting small objects is still difficult because their low-resolution appearance provides insufficient discriminative features, and...
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
MDPI AG
2025-05-01
|
| Series: | AppliedMath |
| Subjects: | |
| Online Access: | https://www.mdpi.com/2673-9909/5/2/58 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Over the past decade, the field of object detection has advanced remarkably, especially in the accurate recognition of medium- and large-sized objects. Nevertheless, detecting small objects is still difficult because their low-resolution appearance provides insufficient discriminative features, and they often suffer severe occlusions, particularly in the safety-critical context of autonomous driving. Conventional detectors often fail to extract sufficient information from shallow feature maps, which limits their ability to detect small objects with high precision. To address this issue, we propose the ECAN-Detector, an efficient context-aggregation method designed to enrich the feature representation of shallow layers, which are particularly beneficial for small-object detection. The model first employs an additional shallow detection layer to extract high-resolution features that provide more detailed information for subsequent stages of the network, and then incorporates a dynamic scaled transformer (DST) that enriches spatial perception by adaptively fusing global semantics and local context. Concurrently, a context-augmentation module (CAM) embedded in the shallow layer complements both global and local features relevant to small objects. To further boost the average precision of small-object detection, we implement a faster method utilizing two reparametrized convolutions in the detection head. Finally, extensive experiments conducted on the VisDrone2012-DET and VisDrone2021-DET datasets verified that our proposed method surpasses the baseline model, and achieved a significant improvement of 3.1% in AP and 3.5% in <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mi>A</mi><msub><mi>P</mi><mi>s</mi></msub></mrow></semantics></math></inline-formula>. Compared with recent state-of-the-art (SOTA) detectors, ECAN Detector delivers comparable accuracy yet preserves real-time throughput, reaching 54.3 FPS. |
|---|---|
| ISSN: | 2673-9909 |