Selected work

Representative Papers

InspectVLM: Unified in Theory, Unreliable in Practice

Aug 03, 2025

This work investigates the feasibility and robustness of unified vision-language models (VLMs) as replacements for task-specific models in industrial inspection. To address diverse requirements—including image classification, object detection, and keypoint localization—it proposes a language-interface-based unified paradigm, introduces InspectMM—the first large-scale multimodal industrial inspection dataset—and performs instruction tuning on Florence-2. Experiments show strong performance on classification and structured keypoint tasks, but limited robustness on fine-grained detection, high sensitivity to prompt engineering, and weaker visual grounding compared to specialized architectures like ResNet. The core contribution lies in empirically exposing the fundamental tension between *linguistic unification* and *visual reliability* in VLMs for industrial applications, thereby establishing an empirical benchmark and identifying concrete directions for advancing VLMs toward high-precision industrial deployment.

0 citationsRead paper

Aerial Infrared Health Monitoring of Solar Photovoltaic Farms at Scale

Mar 03, 2025

Large-scale photovoltaic (PV) power plants often suffer from invisible defects—such as hotspots and connection failures—that hinder accurate operational efficiency assessment. To address this, we propose an end-to-end health monitoring framework tailored for aerial thermal infrared (TIR) remote sensing. Our method introduces the first geographically diverse, kilo-scale aerial TIR dataset covering hundreds of PV plants. We design a thermal anomaly detection pipeline integrating multi-scale processing, georegistration, and weakly supervised segmentation to overcome the limitations of visible-light-based inspection. Additionally, we incorporate a cross-operating-condition robust deployment mechanism. The framework achieves sub-module-level defect localization with >92% accuracy and quantifies associated power loss with <8.5% error. Validated across multiple hundred-megawatt-scale plants in North America, our approach enables scalable, reliable assessment of renewable energy assets.

0 citationsRead paper

Solar Panel Mapping via Oriented Object Detection

Feb 05, 2025

To address the inefficiency and scalability limitations of manual photovoltaic (PV) panel geolocation and orientation mapping in solar farm operations, this paper proposes an end-to-end fine-grained spatial mapping method based on rotated object detection. We introduce rotated bounding box detection—previously unexplored in PV panel geolocation—for the first time, enabling high-precision localization and orientation estimation of arbitrarily oriented panels. Our approach integrates multi-scale feature fusion with an orientation-aware regression module to jointly predict per-panel geographic coordinates and azimuth angles. Evaluated on a large-scale, real-world dataset covering multiple U.S. locations, our method achieves a mean Average Precision (mAP) of 83.3%, significantly outperforming conventional axis-aligned bounding box detectors. This work establishes a scalable, fully automated mapping paradigm for intelligent, large-scale solar farm inspection and digital twin construction.

0 citationsRead paper

Recent publications

Latest Papers

InspectVLM: Unified in Theory, Unreliable in Practice

Aug 03, 2025

This work investigates the feasibility and robustness of unified vision-language models (VLMs) as replacements for task-specific models in industrial inspection. To address diverse requirements—including image classification, object detection, and keypoint localization—it proposes a language-interface-based unified paradigm, introduces InspectMM—the first large-scale multimodal industrial inspection dataset—and performs instruction tuning on Florence-2. Experiments show strong performance on classification and structured keypoint tasks, but limited robustness on fine-grained detection, high sensitivity to prompt engineering, and weaker visual grounding compared to specialized architectures like ResNet. The core contribution lies in empirically exposing the fundamental tension between *linguistic unification* and *visual reliability* in VLMs for industrial applications, thereby establishing an empirical benchmark and identifying concrete directions for advancing VLMs toward high-precision industrial deployment.

0 citationsRead paper

Aerial Infrared Health Monitoring of Solar Photovoltaic Farms at Scale

Mar 03, 2025

Large-scale photovoltaic (PV) power plants often suffer from invisible defects—such as hotspots and connection failures—that hinder accurate operational efficiency assessment. To address this, we propose an end-to-end health monitoring framework tailored for aerial thermal infrared (TIR) remote sensing. Our method introduces the first geographically diverse, kilo-scale aerial TIR dataset covering hundreds of PV plants. We design a thermal anomaly detection pipeline integrating multi-scale processing, georegistration, and weakly supervised segmentation to overcome the limitations of visible-light-based inspection. Additionally, we incorporate a cross-operating-condition robust deployment mechanism. The framework achieves sub-module-level defect localization with >92% accuracy and quantifies associated power loss with <8.5% error. Validated across multiple hundred-megawatt-scale plants in North America, our approach enables scalable, reliable assessment of renewable energy assets.

0 citationsRead paper

Solar Panel Mapping via Oriented Object Detection

Feb 05, 2025

To address the inefficiency and scalability limitations of manual photovoltaic (PV) panel geolocation and orientation mapping in solar farm operations, this paper proposes an end-to-end fine-grained spatial mapping method based on rotated object detection. We introduce rotated bounding box detection—previously unexplored in PV panel geolocation—for the first time, enabling high-precision localization and orientation estimation of arbitrarily oriented panels. Our approach integrates multi-scale feature fusion with an orientation-aware regression module to jointly predict per-panel geographic coordinates and azimuth angles. Evaluated on a large-scale, real-world dataset covering multiple U.S. locations, our method achieves a mean Average Precision (mAP) of 83.3%, significantly outperforming conventional axis-aligned bounding box detectors. This work establishes a scalable, fully automated mapping paradigm for intelligent, large-scale solar farm inspection and digital twin construction.

0 citationsRead paper