Selected work

Representative Papers

Enhancing Generative Information Extraction with Two-step Validation: A Product Attribute Use Case

Jul 29, 2026

This work addresses the challenges of scarce labeled data and weakly expressed, low-salience product attributes in applications such as digital product passports by proposing a two-step verification generative information extraction framework that integrates pretrained language models (PLMs) with large language models (LLMs). The approach first employs a PLM for initial candidate extraction and then leverages a locally deployable open-source LLM—such as those in the Llama family—for secondary verification and error correction, substantially improving extraction accuracy for sparse and weakly expressed entities. Experimental results demonstrate that the proposed framework enhances generalization capability and enables medium-scale models to approach the performance of much larger models, all while preserving data privacy and maintaining computational efficiency. The method has been successfully integrated into a demonstration system tailored for digital product passports.

0 citationsRead paper

Transfer learning-based method for automated ewaste recycling in smart cities

Jun 22, 2026

This study addresses the low efficiency and insufficient accuracy in electronic waste sorting by proposing, for the first time, a transfer learning–based approach tailored to few-shot e-waste classification. Specifically, the authors introduce an intelligent sorting method built upon fine-tuned AlexNet, enhanced with data augmentation, hyperparameter optimization, and an SGD with Momentum optimizer (learning rate 3e-4). Evaluated on a small-scale dataset comprising 12 smartphone categories across six brands, the method achieves a classification accuracy of 98%. The results demonstrate a significant reduction in sorting errors and validate the feasibility of high-precision automated classification under resource-constrained conditions, offering a promising pathway for integrating artificial intelligence into circular economy initiatives.

0 citationsRead paper

Battery detection of XRay images using transfer learning

Jun 10, 2026

This study addresses the challenge of automatic detection, localization, and type identification of lithium-ion batteries in X-ray images by proposing a two-stage transfer learning approach. First, a YOLOv5m model is fine-tuned on a large-scale dataset of electronic devices; subsequently, the adapted model is transferred to the X-ray domain to jointly detect and classify three battery types—prismatic, pouch, and cylindrical. The proposed method achieves a detection accuracy of 94%, representing a 5% improvement over the original pretrained model, while maintaining a real-time inference speed of 22 milliseconds per frame. This approach effectively balances high accuracy with computational efficiency, offering a practical and scalable solution for intelligent battery sorting systems.

0 citationsRead paper

Don't waste SAM

Jun 09, 2026

This study addresses the performance degradation of the general-purpose segmentation model SAM in real-world waste scenarios, where occlusion, deformation, transparency, and background clutter pose significant challenges. For the first time, the authors systematically evaluate and fine-tune SAM’s generalization capability across three waste datasets—Zerowaste, TACO, and TrashCan 1.0. Building upon the SAM-ViT-H architecture and employing a targeted supervised fine-tuning strategy, the approach substantially improves segmentation accuracy on complex waste objects. Experimental results demonstrate that the fine-tuned model achieves IoU gains exceeding 30 on both Zerowaste and TACO, and lags only 1.44 behind the state-of-the-art on TrashCan 1.0, significantly outperforming existing methods overall. These findings validate SAM’s effectiveness and potential as a foundational model for waste segmentation tasks.

0 citationsRead paper

Parameter-Efficient Fine-Tuning of Large Pretrained Models for Instance Segmentation Tasks

Jun 01, 2026

This work addresses the high computational cost of conventional fine-tuning for instance segmentation, which typically requires updating a large fraction (40–55%) of parameters in large pre-trained models. To improve parameter efficiency, the study explores parameter-efficient fine-tuning (PEFT) methods, introducing LoRA into deformable attention mechanisms for the first time and systematically evaluating the trade-offs between performance and efficiency based on the number and placement of adapters within the Transformer architecture. Experimental results demonstrate that by fine-tuning only 1–6% of the model parameters, the proposed approach matches or even surpasses the performance of full fine-tuning across four benchmark datasets. These findings validate the efficacy and feasibility of PEFT for instance segmentation and further reveal that its effectiveness is influenced by dataset complexity and model architecture.

0 citationsRead paper

Recent publications

Latest Papers

Enhancing Generative Information Extraction with Two-step Validation: A Product Attribute Use Case

Jul 29, 2026

This work addresses the challenges of scarce labeled data and weakly expressed, low-salience product attributes in applications such as digital product passports by proposing a two-step verification generative information extraction framework that integrates pretrained language models (PLMs) with large language models (LLMs). The approach first employs a PLM for initial candidate extraction and then leverages a locally deployable open-source LLM—such as those in the Llama family—for secondary verification and error correction, substantially improving extraction accuracy for sparse and weakly expressed entities. Experimental results demonstrate that the proposed framework enhances generalization capability and enables medium-scale models to approach the performance of much larger models, all while preserving data privacy and maintaining computational efficiency. The method has been successfully integrated into a demonstration system tailored for digital product passports.

0 citationsRead paper

Transfer learning-based method for automated ewaste recycling in smart cities

Jun 22, 2026

This study addresses the low efficiency and insufficient accuracy in electronic waste sorting by proposing, for the first time, a transfer learning–based approach tailored to few-shot e-waste classification. Specifically, the authors introduce an intelligent sorting method built upon fine-tuned AlexNet, enhanced with data augmentation, hyperparameter optimization, and an SGD with Momentum optimizer (learning rate 3e-4). Evaluated on a small-scale dataset comprising 12 smartphone categories across six brands, the method achieves a classification accuracy of 98%. The results demonstrate a significant reduction in sorting errors and validate the feasibility of high-precision automated classification under resource-constrained conditions, offering a promising pathway for integrating artificial intelligence into circular economy initiatives.

0 citationsRead paper

Battery detection of XRay images using transfer learning

Jun 10, 2026

This study addresses the challenge of automatic detection, localization, and type identification of lithium-ion batteries in X-ray images by proposing a two-stage transfer learning approach. First, a YOLOv5m model is fine-tuned on a large-scale dataset of electronic devices; subsequently, the adapted model is transferred to the X-ray domain to jointly detect and classify three battery types—prismatic, pouch, and cylindrical. The proposed method achieves a detection accuracy of 94%, representing a 5% improvement over the original pretrained model, while maintaining a real-time inference speed of 22 milliseconds per frame. This approach effectively balances high accuracy with computational efficiency, offering a practical and scalable solution for intelligent battery sorting systems.

0 citationsRead paper

Don't waste SAM

Jun 09, 2026

This study addresses the performance degradation of the general-purpose segmentation model SAM in real-world waste scenarios, where occlusion, deformation, transparency, and background clutter pose significant challenges. For the first time, the authors systematically evaluate and fine-tune SAM’s generalization capability across three waste datasets—Zerowaste, TACO, and TrashCan 1.0. Building upon the SAM-ViT-H architecture and employing a targeted supervised fine-tuning strategy, the approach substantially improves segmentation accuracy on complex waste objects. Experimental results demonstrate that the fine-tuned model achieves IoU gains exceeding 30 on both Zerowaste and TACO, and lags only 1.44 behind the state-of-the-art on TrashCan 1.0, significantly outperforming existing methods overall. These findings validate SAM’s effectiveness and potential as a foundational model for waste segmentation tasks.

0 citationsRead paper

Parameter-Efficient Fine-Tuning of Large Pretrained Models for Instance Segmentation Tasks

Jun 01, 2026

This work addresses the high computational cost of conventional fine-tuning for instance segmentation, which typically requires updating a large fraction (40–55%) of parameters in large pre-trained models. To improve parameter efficiency, the study explores parameter-efficient fine-tuning (PEFT) methods, introducing LoRA into deformable attention mechanisms for the first time and systematically evaluating the trade-offs between performance and efficiency based on the number and placement of adapters within the Transformer architecture. Experimental results demonstrate that by fine-tuning only 1–6% of the model parameters, the proposed approach matches or even surpasses the performance of full fine-tuning across four benchmark datasets. These findings validate the efficacy and feasibility of PEFT for instance segmentation and further reveal that its effectiveness is influenced by dataset complexity and model architecture.

0 citationsRead paper