Selected work

Representative Papers

Applying Deep Learning for cockpit segmentation in the context of mixed reality

Jun 02, 2026

This study addresses the demand for precise foreground-background segmentation in mixed reality (MR) applications involving heavy machinery cockpits, where seamless integration of virtual and real elements is critical. For the first time, U-Net and DeepLabV3+ are applied to semantic segmentation of first-person cockpit imagery from a mining truck simulator. Trained on a dataset collected from real-world scenarios, both models achieve approximately 90% segmentation accuracy while maintaining real-time performance. Experimental results demonstrate that the selected architectures effectively and accurately delineate cockpit foregrounds from backgrounds, significantly enhancing immersion and compositing quality in MR environments. This work thus provides robust foundational visual support for MR systems deployed with heavy equipment.

0 citationsRead paper

Mathematical Morphology in Machine Learning

May 28, 2026

This study addresses the limitations of traditional machine learning methods in capturing the shape, density, and fractal structure of complex data distributions. It pioneers the systematic integration of mathematical morphology into machine learning by introducing a fast clustering algorithm based on morphological reconstruction, which inherently possesses maximum-cluster awareness and cost-free denoising capabilities. Furthermore, the authors propose a novel hybrid distance metric combining Minkowski and Chebyshev distances, demonstrating significantly higher computational efficiency in the discrete space ℤ²—1.3× faster than Manhattan and 329.5× faster than Euclidean distance. Experimental results across 33 UCI datasets show that a k-nearest neighbor classifier equipped with this metric surpasses average accuracy on 26 datasets and achieves state-of-the-art performance on 9, confirming the method’s superior ability to jointly model geometric and topological characteristics of data.

0 citationsRead paper

A Novel Approach for the Counting of Wood Logs Using cGANs and Image Processing Techniques

May 22, 2026

This study addresses the challenge of accurately counting eucalyptus logs in complex scenarios involving occlusion, overlap, and cluttered backgrounds. To this end, it introduces conditional generative adversarial networks (cGANs) for the first time to this task, complemented by tailored image pre- and post-processing strategies that effectively mitigate noise and adhesion artifacts. Log counts are derived through connected component analysis, enabling efficient and robust enumeration. The authors construct and publicly release a large-scale dataset comprising 466 images. Evaluated on an NVIDIA T4 GPU, the proposed method achieves real-time performance with an average processing time of 0.713 seconds per image. Experimental results demonstrate a pixel-level accuracy of 96.4%, a log counting accuracy of 92.3%, F1 scores ranging from 0.879 to 0.933, and Intersection over Union (IoU) values between 0.784 and 0.875.

0 citationsRead paper

Machine learning applied to emerald gemstone grading: framework proposal and creation of a public dataset

May 22, 2026

This study addresses the high subjectivity and inconsistency in emerald grading caused by manual comparison against reference stones. To overcome these limitations, the work proposes the first automated grading framework that integrates image processing with conventional (non-deep learning) machine learning techniques. The system requires only minimal human intervention—placing the gemstone into an imaging chamber—and then autonomously handles image acquisition through to final classification. Key contributions include the creation and public release of the first open dataset comprising 192 emerald images along with pre-extracted features, and the achievement of a 98% grading accuracy, which significantly outperforms existing deep learning approaches. This advancement markedly enhances the objectivity and reproducibility of emerald quality assessment.

0 citationsRead paper

Local-sensitive connectivity filter (ls-cf): A post-processing unsupervised improvement of the frangi, hessian and vesselness filters for multimodal vessel segmentation

May 20, 2026

This study addresses the challenge of discontinuous vessel responses in retinal vascular segmentation, where traditional filters such as Frangi’s often produce fragmented vessels, thereby compromising accurate extraction of vascular structures across multimodal medical images. To overcome this limitation, the authors propose LS-CF, an unsupervised post-processing method that, for the first time, integrates local sensitivity–aware connectivity constraints with a heuristic tolerance mechanism to evaluate and reconnect broken vessel segments at the pixel level—without requiring any training data, thus enabling generalization across diverse imaging modalities. Built upon Frangi responses, LS-CF combines local connectivity analysis with morphological operations to form a lightweight yet effective unsupervised filter. The method achieves state-of-the-art performance on multiple benchmark datasets—including OSIRIX, IOSTAR, DRIVE, STARE, and CHASE_DB—with particularly notable gains over existing unsupervised approaches on CHASE_DB.

0 citationsRead paper

Recent publications

Latest Papers

Applying Deep Learning for cockpit segmentation in the context of mixed reality

Jun 02, 2026

This study addresses the demand for precise foreground-background segmentation in mixed reality (MR) applications involving heavy machinery cockpits, where seamless integration of virtual and real elements is critical. For the first time, U-Net and DeepLabV3+ are applied to semantic segmentation of first-person cockpit imagery from a mining truck simulator. Trained on a dataset collected from real-world scenarios, both models achieve approximately 90% segmentation accuracy while maintaining real-time performance. Experimental results demonstrate that the selected architectures effectively and accurately delineate cockpit foregrounds from backgrounds, significantly enhancing immersion and compositing quality in MR environments. This work thus provides robust foundational visual support for MR systems deployed with heavy equipment.

0 citationsRead paper

Mathematical Morphology in Machine Learning

May 28, 2026

This study addresses the limitations of traditional machine learning methods in capturing the shape, density, and fractal structure of complex data distributions. It pioneers the systematic integration of mathematical morphology into machine learning by introducing a fast clustering algorithm based on morphological reconstruction, which inherently possesses maximum-cluster awareness and cost-free denoising capabilities. Furthermore, the authors propose a novel hybrid distance metric combining Minkowski and Chebyshev distances, demonstrating significantly higher computational efficiency in the discrete space ℤ²—1.3× faster than Manhattan and 329.5× faster than Euclidean distance. Experimental results across 33 UCI datasets show that a k-nearest neighbor classifier equipped with this metric surpasses average accuracy on 26 datasets and achieves state-of-the-art performance on 9, confirming the method’s superior ability to jointly model geometric and topological characteristics of data.

0 citationsRead paper

A Novel Approach for the Counting of Wood Logs Using cGANs and Image Processing Techniques

May 22, 2026

This study addresses the challenge of accurately counting eucalyptus logs in complex scenarios involving occlusion, overlap, and cluttered backgrounds. To this end, it introduces conditional generative adversarial networks (cGANs) for the first time to this task, complemented by tailored image pre- and post-processing strategies that effectively mitigate noise and adhesion artifacts. Log counts are derived through connected component analysis, enabling efficient and robust enumeration. The authors construct and publicly release a large-scale dataset comprising 466 images. Evaluated on an NVIDIA T4 GPU, the proposed method achieves real-time performance with an average processing time of 0.713 seconds per image. Experimental results demonstrate a pixel-level accuracy of 96.4%, a log counting accuracy of 92.3%, F1 scores ranging from 0.879 to 0.933, and Intersection over Union (IoU) values between 0.784 and 0.875.

0 citationsRead paper

Machine learning applied to emerald gemstone grading: framework proposal and creation of a public dataset

May 22, 2026

This study addresses the high subjectivity and inconsistency in emerald grading caused by manual comparison against reference stones. To overcome these limitations, the work proposes the first automated grading framework that integrates image processing with conventional (non-deep learning) machine learning techniques. The system requires only minimal human intervention—placing the gemstone into an imaging chamber—and then autonomously handles image acquisition through to final classification. Key contributions include the creation and public release of the first open dataset comprising 192 emerald images along with pre-extracted features, and the achievement of a 98% grading accuracy, which significantly outperforms existing deep learning approaches. This advancement markedly enhances the objectivity and reproducibility of emerald quality assessment.

0 citationsRead paper

Local-sensitive connectivity filter (ls-cf): A post-processing unsupervised improvement of the frangi, hessian and vesselness filters for multimodal vessel segmentation

May 20, 2026

This study addresses the challenge of discontinuous vessel responses in retinal vascular segmentation, where traditional filters such as Frangi’s often produce fragmented vessels, thereby compromising accurate extraction of vascular structures across multimodal medical images. To overcome this limitation, the authors propose LS-CF, an unsupervised post-processing method that, for the first time, integrates local sensitivity–aware connectivity constraints with a heuristic tolerance mechanism to evaluate and reconnect broken vessel segments at the pixel level—without requiring any training data, thus enabling generalization across diverse imaging modalities. Built upon Frangi responses, LS-CF combines local connectivity analysis with morphological operations to form a lightweight yet effective unsupervised filter. The method achieves state-of-the-art performance on multiple benchmark datasets—including OSIRIX, IOSTAR, DRIVE, STARE, and CHASE_DB—with particularly notable gains over existing unsupervised approaches on CHASE_DB.

0 citationsRead paper