Xiangyang Xu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Xiangyang Xu has a total of 7 co-authors. Here are the top co-authors by number of shared publications:
- Haoran Wei
State Key Laboratory of Intelligent Control and Decision of Complex Systems, Beijing Institute of Technology, Beijing 100081, China
1 shared publication - Kailiang Chen
School of Automation, Beijing Institute of Technology, Beijing 100081, China
1 shared publication - Ni Ou
State Key Laboratory of Intelligent Control and Decision of Complex Systems, Beijing Institute of Technology, Beijing 100081, China
1 shared publication - Qian Li
School of Automation, Beijing Institute of Technology, Beijing 100081, China
1 shared publication - Xinru Zhang
School of Integrated Circuits and Electronics, Beijing Institute of Technology, Beijing, China
1 shared publication - Yaping Dai
School of Automation, Beijing Institute of Technology, Beijing 100081, China
1 shared publication
Latest Publications
- Open Access
- 8 July 2026
- Published by MDPI AG in Remote Sensing
- Vol. 18 (14) , 2288
Crop-Tool-Augmented Active Perception with Reinforcement Learning for High-Resolution Remote Sensing Visual Question Answering
High-resolution remote sensing visual question answering (RS-VQA) requires models to identify question-relevant regions and reason over fine-grained visual evidence. However, existing vision–language models usually rely on fixed global image inputs, which may lose critical local details in ultra-high-resolution imagery and struggle with sparse informative regions, large object-scale variations, and complex spatial layouts. To address these challenges, this paper proposes a crop-tool-augmented active perception framework with reinforcement learning. The framework introduces structured tokens to explicitly organize the reasoning process into question understanding, cropping decision-making, local evidence acquisition and final answer generation. Based on this design, the model can actively determine whether a cropping operation is needed and select task-relevant regions for further inspection. To enable stable tool-use and multi-turn reasoning in a compact vision–language model, we construct teacher-guided cropping reasoning trajectories from high-resolution images, question–answer pairs, and annotated regions in the LRS-GRO dataset, and use them for cold-start supervised fine-tuning of Qwen2.5-VL-3B. Furthermore, we introduce Group Relative Policy Optimization to refine the model’s active perception policy. A region-aware reward function is designed by integrating output-format constraints, reference-region coverage, answer semantic consistency, and cropping penalties, which encourages compact and informative region selection while reducing redundant tool invocations. Experiments on VRSBench, MME-RealWorld-RS, XLRS-Bench, and LRS-VQA demonstrate that the proposed method achieves competitive overall performance compared with closed-source, open-source, and remote-sensing-specific vision–language models, and obtains the best or comparable results on most benchmarks. Ablation studies further verify the effectiveness of structured supervised fine-tuning, reinforcement learning optimization, and the proposed reward design.
- 8 July 2026
- Published by MDPI AG in Remote Sensing
- Vol. 18 (14) , 2288
Open AccessCrop-Tool-Augmented Active Perception with Reinforcement Learning for High-Resolution Remote Sensing Visual Question Answering
High-resolution remote sensing visual question answering (RS-VQA) requires models to identify question-relevant regions and reason over fine-grained visual evidence. However, existing vision–language models usually rely on fixed global image inputs, which may lose critical local details in ultra-high-resolution imagery and struggle with sparse informative regions, large object-scale variations, and complex spatial layouts. To address these challenges, this paper proposes a crop-tool-augmented active perception framework with reinforcement learning. The framework introduces structured tokens to explicitly organize the reasoning process into question understanding, cropping decision-making, local evidence acquisition and final answer generation. Based on this design, the model can actively determine whether a cropping operation is needed and select task-relevant regions for further inspection. To enable stable tool-use and multi-turn reasoning in a compact vision–language model, we construct teacher-guided cropping reasoning trajectories from high-resolution images, question–answer pairs, and annotated regions in the LRS-GRO dataset, and use them for cold-start supervised fine-tuning of Qwen2.5-VL-3B. Furthermore, we introduce Group Relative Policy Optimization to refine the model’s active perception policy. A region-aware reward function is designed by integrating output-format constraints, reference-region coverage, answer semantic consistency, and cropping penalties, which encourages compact and informative region selection while reducing redundant tool invocations. Experiments on VRSBench, MME-RealWorld-RS, XLRS-Bench, and LRS-VQA demonstrate that the proposed method achieves competitive overall performance compared with closed-source, open-source, and remote-sensing-specific vision–language models, and obtains the best or comparable results on most benchmarks. Ablation studies further verify the effectiveness of structured supervised fine-tuning, reinforcement learning optimization, and the proposed reward design.
- Open Access
- 29 September 2021
- Published by MDPI AG in Remote Sensing
- Vol. 13 (19) , 3900
DEANet: Dual Encoder with Attention Network for Semantic Segmentation of Remote Sensing Imagery
Remote sensing has now been widely used in various fields, and the research on the automatic land-cover segmentation methods of remote sensing imagery is significant to the development of remote sensing technology. Deep learning methods, which are developing rapidly in the field of semantic segmentation, have been widely applied to remote sensing imagery segmentation. In this work, a novel deep learning network—Dual Encoder with Attention Network (DEANet) is proposed. In this network, a dual-branch encoder structure, whose first branch is used to generate a rough guidance feature map as area attention to help re-encode feature maps in the next branch, is proposed to improve the encoding ability of the network, and an improved pyramid partial decoder (PPD) based on the parallel partial decoder is put forward to make fuller use of the features form the encoder along with the receptive filed block (RFB). In addition, an edge attention module using the transfer learning method is introduced to explicitly advance the segmentation performance in edge areas. Except for structure, a loss function composed with the weighted Cross Entropy (CE) loss and weighted Union subtract Intersection (UsI) loss is designed for training, where UsI loss represents a new region-based aware loss which replaces the IoU loss to adapt to multi-classification tasks. Furthermore, a detailed training strategy for the network is introduced as well. Extensive experiments on three public datasets verify the effectiveness of each proposed module in our framework and demonstrate that our method achieves more excellent performance over some state-of-the-art methods.
- 29 September 2021
- Published by MDPI AG in Remote Sensing
- Vol. 13 (19) , 3900
Open AccessDEANet: Dual Encoder with Attention Network for Semantic Segmentation of Remote Sensing Imagery
Remote sensing has now been widely used in various fields, and the research on the automatic land-cover segmentation methods of remote sensing imagery is significant to the development of remote sensing technology. Deep learning methods, which are developing rapidly in the field of semantic segmentation, have been widely applied to remote sensing imagery segmentation. In this work, a novel deep learning network—Dual Encoder with Attention Network (DEANet) is proposed. In this network, a dual-branch encoder structure, whose first branch is used to generate a rough guidance feature map as area attention to help re-encode feature maps in the next branch, is proposed to improve the encoding ability of the network, and an improved pyramid partial decoder (PPD) based on the parallel partial decoder is put forward to make fuller use of the features form the encoder along with the receptive filed block (RFB). In addition, an edge attention module using the transfer learning method is introduced to explicitly advance the segmentation performance in edge areas. Except for structure, a loss function composed with the weighted Cross Entropy (CE) loss and weighted Union subtract Intersection (UsI) loss is designed for training, where UsI loss represents a new region-based aware loss which replaces the IoU loss to adapt to multi-classification tasks. Furthermore, a detailed training strategy for the network is introduced as well. Extensive experiments on three public datasets verify the effectiveness of each proposed module in our framework and demonstrate that our method achieves more excellent performance over some state-of-the-art methods.