An Investigation of Infrared Small Target Detection by Using the SPT–YOLO Technique

Technologies, Journal Year: 2025, Volume and Issue: 13(1), P. 40 - 40

Published: Jan. 17, 2025

To detect and recognize small-size submerged complex background targets in infrared images, we combine a dynamic receptive field fusion strategy multi-scale feature mechanism to improve the detection performance of small significantly. The space-to-depth convolution module is introduced as downsampling layer backbone first achieves same sampling effect. More detailed information retained at time. Thus, model’s capability for has been enhanced. Then, pyramid level 2 map with minimum maximum resolution added neck, which reduces loss positional during sampling. Furthermore, x-small heads are added, understanding overall characteristics structure target enhanced much more, representation localization have improved. Finally, cross-entropy function original network model replaced by an adaptive threshold focal function, forcing allocate more attention features. above methods based on public tool, eighth version You Only Look Once (YOLO) improved, it named SPT–YOLO (SPDConv + P2 Adaptive Threshold YOLOV8s) this paper. Some experiments datasets such object (IR-SOD) 1K(IRSTD-1K), etc. executed verify proposed algorithm; mean average precision 94.0% 69% under condition 0.5 over range from 0.95 obtained, respectively. results show that method best compared existing methods.

Language: Английский

TTMGNet: Tree Topology Mamba-Guided Network Collaborative Hierarchical Incremental Aggregation for Change Detection

Remote Sensing, Journal Year: 2024, Volume and Issue: 16(21), P. 4068 - 4068

Published: Oct. 31, 2024

Change detection (CD) identifies surface changes by analyzing bi-temporal remote sensing (RS) images of the same region and is essential for effective urban planning, ensuring optimal allocation resources, supporting disaster management efforts. However, deep-learning-based CD methods struggle with background noise pseudo-changes due to local receptive field limitations or computing resource constraints, which limits long-range dependency capture feature integration, normally resulting in fragmented detections high false positive rates. To address these challenges, we propose a tree topology Mamba-guided network (TTMGNet) based on Mamba architecture, combines architecture effectively capturing global features, unique structure retaining fine details, hierarchical fusion mechanism that enhances multi-scale integration robustness against noise. Specifically, Tree Topology Feature Extractor (TTMFE) leverages similarity pixels generate minimum spanning (MST) sequences, guiding information aggregation transmission. This approach utilizes State Space Model (TTSSM) embed spatial positional while preserving extraction capability, thereby features. Subsequently, Hierarchical Incremental Aggregation Module utilized gradually align merge features from deep shallow layers facilitate integration. Through residual connections cross-channel attention (CCA), HIAM interaction between neighboring maps, critical are retained during process, enabling more accurate results CD. The proposed TTMGNet achieved F1 scores 92.31% LEVIR-CD, 90.94% WHU-CD, 77.25% CL-CD, outperforming current mainstream suppressing impact pseudo-change accurately identifying change regions.

Language: Английский

Remote Sensing Identification of Picea schrenkiana var. tianschanica in GF-1 Images Based on a Multiple Mixed Attention U-Net Model

Forests, Journal Year: 2024, Volume and Issue: 15(11), P. 2039 - 2039

Published: Nov. 19, 2024

Remote sensing technology plays an important role in woodland identification. However, mountainous areas with complex terrain, accurate extraction of boundary information still faces challenges. To address this problem, paper proposes a multiple mixed attention U-Net (MMA-U-Net) semantic segmentation model using 2015 and 2022 GF-1 PMS images as data sources to improve the ability extract features Picea schrenkiana var. tianschanica forest. The architecture serves its underlying network, feature is improved by adding hybrid CBAM replacing original skip connection DCA module accuracy segmentation. results show that on remote dataset images, compared other models, increased 5.42%–19.84%. By statistically analyzing spatial distribution well their changes, area was 3471.38 km2 3726.10 2022. Combining predicted DEM data, it found were most distributed at altitude 1700–2500 m. method proposed study can accurately identify provides theoretical basis research direction for forest monitoring.

Language: Английский