Selected work

Representative Papers

Thermodynamic Regulation of Finite-Time Gibbs Training in Energy-Based Models: A Restricted Boltzmann Machine Study

Mar 02, 2026

This work addresses the challenges of sampling collapse, negative-phase localization, and parameter drift in Restricted Boltzmann Machine (RBM) training under fixed-temperature finite-time Gibbs sampling. To overcome these issues, the authors propose an endogenous thermodynamic regulation framework that treats temperature as a dynamic state variable coupled with sampling statistics, thereby modeling RBM training as a controlled non-equilibrium dynamical process. By incorporating a thermodynamic self-regulation mechanism, the approach rigorously ensures global parameter boundedness under L2 regularization and local exponential stability of subsystems, effectively preventing inverse-temperature explosion and sampling freeze. Theoretical analysis leverages two-time-scale separation and local Lipschitz conditions. Experiments on MNIST demonstrate significantly improved normalized stability and effective sample size while preserving reconstruction performance, outperforming fixed-temperature baselines.

0 citationsRead paper

A Novel Public Dataset for Strawberry (Fragaria x ananassa) Ripeness Detection and Comparative Evaluation of YOLO-Based Models

Feb 17, 2026

Strawberry maturity assessment has long relied on subjective visual inspection, and the absence of a standardized public dataset has hindered algorithmic development. To address this gap, this work presents the first publicly available dataset for strawberry maturity detection, comprising 566 images with 1,201 meticulously annotated instances captured under varying lighting conditions and greenhouse environments. Leveraging this dataset, the study systematically evaluates models from the YOLOv8, YOLOv9, and YOLO11 families. Experimental results show that YOLOv9c achieves the highest precision (90.94%), YOLO11s attains the best recall (83.74%), and YOLOv8s delivers the strongest overall performance with an mAP@50 of 86.09%. These findings underscore the efficacy of small-to-medium-scale models in agricultural vision tasks and establish a reliable benchmark for intelligent harvesting systems.

0 citationsRead paper

A Mobile Application for Flower Recognition System Based on Convolutional Neural Networks

Jan 22, 2026

This study addresses the challenge faced by non-experts in accurately identifying flower species in the wild, a task traditionally reliant on expert consultation that suffers from limited accessibility. To overcome this, the authors develop a mobile-based flower recognition system leveraging lightweight convolutional neural networks. They systematically evaluate three architectures—MobileNet, DenseNet121, and Xception—combined with various optimization algorithms. Experimental results demonstrate that DenseNet121 trained with the SGD optimizer achieves the best performance, yielding an accuracy of 95.84% on a custom dataset, with precision, recall, and F1-score all reaching 96.00%. The resulting system enables high-accuracy, low-latency real-time flower identification, offering non-specialist users an efficient and accessible tool for plant recognition in natural settings.

0 citationsRead paper

FruitProm: Probabilistic Maturity Estimation and Detection of Fruits and Vegetables

Oct 28, 2025

Existing methods formulate fruit and vegetable ripeness estimation as discrete classification, contradicting its inherently continuous evolutionary nature and thereby introducing boundary ambiguity and information loss. To address this, we propose the first continuous probabilistic learning framework for ripeness estimation, reformulating RT-DETRv2 as a probabilistic detector. Our approach introduces a learnable distribution-output head that jointly optimizes object localization, ripeness-level classification, and uncertainty modeling. Specifically, it directly predicts a continuous probability distribution—parameterized by mean and variance—for ripeness, ensuring biological plausibility while enabling confidence-aware decision-making for robotic applications. Evaluated on a large-scale fruit-and-vegetable dataset, our method achieves 85.6% mAP, significantly improving both fine-grained ripeness-state assessment accuracy and prediction reliability compared to discrete alternatives.

0 citationsRead paper

CoAtNeXt:An Attention-Enhanced ConvNeXtV2-Transformer Hybrid Model for Gastric Tissue Classification

Sep 11, 2025

To address the low efficiency, poor inter-observer consistency, and insufficient standardization in manual interpretation of gastric histopathological images, this paper proposes an attention-enhanced hybrid ConvNeXtV2-Transformer model. Methodologically, we replace the MBConv blocks in CoAtNet with ConvNeXtV2 modules and integrate the CBAM channel-spatial joint attention mechanism, thereby preserving fine-grained local feature representation while enhancing global contextual modeling. The resulting end-to-end classification framework achieves 96.47% and 98.29% accuracy on the HMU-GC-HE-30K and GasHisSDB datasets, respectively, with a peak AUC of 99.90%. These results significantly surpass those of state-of-the-art CNNs and Vision Transformers. Our approach provides a highly accurate and robust solution for automated early diagnosis of gastric diseases.

0 citationsRead paper

Recent publications

Latest Papers

Thermodynamic Regulation of Finite-Time Gibbs Training in Energy-Based Models: A Restricted Boltzmann Machine Study

Mar 02, 2026

This work addresses the challenges of sampling collapse, negative-phase localization, and parameter drift in Restricted Boltzmann Machine (RBM) training under fixed-temperature finite-time Gibbs sampling. To overcome these issues, the authors propose an endogenous thermodynamic regulation framework that treats temperature as a dynamic state variable coupled with sampling statistics, thereby modeling RBM training as a controlled non-equilibrium dynamical process. By incorporating a thermodynamic self-regulation mechanism, the approach rigorously ensures global parameter boundedness under L2 regularization and local exponential stability of subsystems, effectively preventing inverse-temperature explosion and sampling freeze. Theoretical analysis leverages two-time-scale separation and local Lipschitz conditions. Experiments on MNIST demonstrate significantly improved normalized stability and effective sample size while preserving reconstruction performance, outperforming fixed-temperature baselines.

0 citationsRead paper

A Novel Public Dataset for Strawberry (Fragaria x ananassa) Ripeness Detection and Comparative Evaluation of YOLO-Based Models

Feb 17, 2026

Strawberry maturity assessment has long relied on subjective visual inspection, and the absence of a standardized public dataset has hindered algorithmic development. To address this gap, this work presents the first publicly available dataset for strawberry maturity detection, comprising 566 images with 1,201 meticulously annotated instances captured under varying lighting conditions and greenhouse environments. Leveraging this dataset, the study systematically evaluates models from the YOLOv8, YOLOv9, and YOLO11 families. Experimental results show that YOLOv9c achieves the highest precision (90.94%), YOLO11s attains the best recall (83.74%), and YOLOv8s delivers the strongest overall performance with an mAP@50 of 86.09%. These findings underscore the efficacy of small-to-medium-scale models in agricultural vision tasks and establish a reliable benchmark for intelligent harvesting systems.

0 citationsRead paper

A Mobile Application for Flower Recognition System Based on Convolutional Neural Networks

Jan 22, 2026

This study addresses the challenge faced by non-experts in accurately identifying flower species in the wild, a task traditionally reliant on expert consultation that suffers from limited accessibility. To overcome this, the authors develop a mobile-based flower recognition system leveraging lightweight convolutional neural networks. They systematically evaluate three architectures—MobileNet, DenseNet121, and Xception—combined with various optimization algorithms. Experimental results demonstrate that DenseNet121 trained with the SGD optimizer achieves the best performance, yielding an accuracy of 95.84% on a custom dataset, with precision, recall, and F1-score all reaching 96.00%. The resulting system enables high-accuracy, low-latency real-time flower identification, offering non-specialist users an efficient and accessible tool for plant recognition in natural settings.

0 citationsRead paper

FruitProm: Probabilistic Maturity Estimation and Detection of Fruits and Vegetables

Oct 28, 2025

Existing methods formulate fruit and vegetable ripeness estimation as discrete classification, contradicting its inherently continuous evolutionary nature and thereby introducing boundary ambiguity and information loss. To address this, we propose the first continuous probabilistic learning framework for ripeness estimation, reformulating RT-DETRv2 as a probabilistic detector. Our approach introduces a learnable distribution-output head that jointly optimizes object localization, ripeness-level classification, and uncertainty modeling. Specifically, it directly predicts a continuous probability distribution—parameterized by mean and variance—for ripeness, ensuring biological plausibility while enabling confidence-aware decision-making for robotic applications. Evaluated on a large-scale fruit-and-vegetable dataset, our method achieves 85.6% mAP, significantly improving both fine-grained ripeness-state assessment accuracy and prediction reliability compared to discrete alternatives.

0 citationsRead paper

CoAtNeXt:An Attention-Enhanced ConvNeXtV2-Transformer Hybrid Model for Gastric Tissue Classification

Sep 11, 2025

To address the low efficiency, poor inter-observer consistency, and insufficient standardization in manual interpretation of gastric histopathological images, this paper proposes an attention-enhanced hybrid ConvNeXtV2-Transformer model. Methodologically, we replace the MBConv blocks in CoAtNet with ConvNeXtV2 modules and integrate the CBAM channel-spatial joint attention mechanism, thereby preserving fine-grained local feature representation while enhancing global contextual modeling. The resulting end-to-end classification framework achieves 96.47% and 98.29% accuracy on the HMU-GC-HE-30K and GasHisSDB datasets, respectively, with a peak AUC of 99.90%. These results significantly surpass those of state-of-the-art CNNs and Vision Transformers. Our approach provides a highly accurate and robust solution for automated early diagnosis of gastric diseases.

0 citationsRead paper