Infrared Tiny Structureless Object Detection Enhanced by Video Super-Resolution
The precise detection of infrared (IR) tiny objects against complex backgrounds is of great significance in the field of aircraft imaging guidance. However, due to the extremely small target features, low brightness, and the tendency to be submerged in complex backgrounds, the detection of IR tiny a...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/11007527/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The precise detection of infrared (IR) tiny objects against complex backgrounds is of great significance in the field of aircraft imaging guidance. However, due to the extremely small target features, low brightness, and the tendency to be submerged in complex backgrounds, the detection of IR tiny aerial targets remains a challenge. Concurrently, super-resolution technology is a novel technique that has emerged in recent years, aiming to enhance image quality without altering hardware specifications, thereby magnifying the characteristics of tiny targets. In light of this, we introduce the infrared tiny object detection method enhanced by video super-resolution. Our proposed method is a two-stage model. Initially, it enhances target features through bidirectional propagation and optical flow alignment within the super-resolution module. Subsequently, it employs restricted receptive field convolution and multi-receptive field feature fusion to enhance the detection accuracy of small targets. Experimental results on the SIATD data set show that our method performs better state-of-the-art approaches, with an F1 score of 0.957, precision of 99.2%, and recall of 92.4%. Ablation studies confirm that the super-resolution module contributes to 35% of the performance, while the restricted receptive field design accounts for another 28% improvement. Furthermore, our model maintains sub-real-time performance at 50ms per frame with only 0.37M parameters, making it suitable for deployment in resource-constrained environments. |
|---|---|
| ISSN: | 2169-3536 |