Selected work

Representative Papers

Detection and Classification of Diseases in Multi-Crop Leaves using LSTM and CNN Models

Mar 01, 2025Journal of Innovative Image Processing

Addressing the challenges of early diagnosis and poor cross-crop generalization in multi-crop leaf disease identification, this study constructs and comparatively evaluates two end-to-end deep learning architectures—CNN and LSTM—using a large-scale dataset of 72,000 leaf images spanning multiple crop species. It presents the first systematic performance assessment of both models on a fine-grained, 38-class multi-crop disease recognition task. Models are trained with data augmentation, the Adam optimizer, and categorical cross-entropy loss. Experimental results show that the CNN achieves a validation accuracy of 96.4%, significantly outperforming the LSTM (93.43%); all key metrics—including F1-score—meet stringent evaluation criteria. This work demonstrates that CNNs exhibit superior fine-grained discriminative capability and cross-crop generalization for plant disease recognition. Moreover, the optimized CNN model is lightweight and robust, exhibiting strong potential for real-world field deployment in precision agriculture applications.

0 citationsRead paper

Underwater Image Enhancement using Generative Adversarial Networks: A Survey

Jan 10, 2025

Underwater images suffer from severe degradation—including blurriness, low contrast, and chromatic distortion—due to light attenuation, scattering, and wavelength-dependent absorption, hindering applications in marine ecological monitoring, underwater archaeology, and AUV navigation. This paper presents the first systematic survey of GAN-based underwater image enhancement methods, covering physics-informed modeling, CNN-GAN hybrid architectures (e.g., U-Net+GAN, CycleGAN variants), multi-scale feature fusion, and perception-driven loss design. We propose a unified evaluation framework integrating benchmark datasets (UIEB, EUVP) and quantitative metrics (UCIQE, UIQM), revealing three critical bottlenecks: poor generalizability, high computational overhead, and dataset bias. To address these, we introduce two novel technical directions: (i) interpretable physics-guided prior embedding and (ii) lightweight co-optimization. Our work establishes a standardized benchmark and provides a comprehensive research roadmap for future advances in underwater image enhancement.

0 citationsRead paper

Recent publications

Latest Papers

Detection and Classification of Diseases in Multi-Crop Leaves using LSTM and CNN Models

Mar 01, 2025Journal of Innovative Image Processing

Addressing the challenges of early diagnosis and poor cross-crop generalization in multi-crop leaf disease identification, this study constructs and comparatively evaluates two end-to-end deep learning architectures—CNN and LSTM—using a large-scale dataset of 72,000 leaf images spanning multiple crop species. It presents the first systematic performance assessment of both models on a fine-grained, 38-class multi-crop disease recognition task. Models are trained with data augmentation, the Adam optimizer, and categorical cross-entropy loss. Experimental results show that the CNN achieves a validation accuracy of 96.4%, significantly outperforming the LSTM (93.43%); all key metrics—including F1-score—meet stringent evaluation criteria. This work demonstrates that CNNs exhibit superior fine-grained discriminative capability and cross-crop generalization for plant disease recognition. Moreover, the optimized CNN model is lightweight and robust, exhibiting strong potential for real-world field deployment in precision agriculture applications.

0 citationsRead paper

Underwater Image Enhancement using Generative Adversarial Networks: A Survey

Jan 10, 2025

Underwater images suffer from severe degradation—including blurriness, low contrast, and chromatic distortion—due to light attenuation, scattering, and wavelength-dependent absorption, hindering applications in marine ecological monitoring, underwater archaeology, and AUV navigation. This paper presents the first systematic survey of GAN-based underwater image enhancement methods, covering physics-informed modeling, CNN-GAN hybrid architectures (e.g., U-Net+GAN, CycleGAN variants), multi-scale feature fusion, and perception-driven loss design. We propose a unified evaluation framework integrating benchmark datasets (UIEB, EUVP) and quantitative metrics (UCIQE, UIQM), revealing three critical bottlenecks: poor generalizability, high computational overhead, and dataset bias. To address these, we introduce two novel technical directions: (i) interpretable physics-guided prior embedding and (ii) lightweight co-optimization. Our work establishes a standardized benchmark and provides a comprehensive research roadmap for future advances in underwater image enhancement.

0 citationsRead paper