ID-UNet: A densely connected UNet architecture for infrared small target segmentation
Existing CNN-based approaches face challenges in effectively and efficiently managing diverse scales of small infrared objects within intricate scenes, primarily as a result of the aggregation effect induced by pooling layers. As a consequence, crucial deep targets may be lost. To tackle this challe...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-01-01
|
| Series: | Alexandria Engineering Journal |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S1110016824011323 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Existing CNN-based approaches face challenges in effectively and efficiently managing diverse scales of small infrared objects within intricate scenes, primarily as a result of the aggregation effect induced by pooling layers. As a consequence, crucial deep targets may be lost. To tackle this challenge, This research proposes an infrared deep dense connection network, termed ID-UNet. Specifically, this research devises a feature extraction module, named Infrared Small Target Feature Extraction (ISTFE), that is embedded within the ID-UNet architecture to enable cross-layer and continuous interaction between deep high-level and shallow low-level features. Consecutive connections within ISTFE’s fusion facilitate the preservation of semantic information for infrared small targets in deep layers, as well as the resolution information in shallow layers. Additionally, the UNet structure parameters were compressed, reducing the parameters by 81% compared to the traditional UNet configuration. Upon evaluating the proposed technique on three typical public datasets, the results demonstrate that the proposed method surpasses all other methods in segmentation metrics, including Intersection over Union (IoU), normalized IoU (nIoU), and F1 score. The proposed method achieves a double-win between high-precision segmentation and low computation requirements. The code is available from https://github.com/AngryWaves/ID-UNet. |
|---|---|
| ISSN: | 1110-0168 |