Meng'en Qin (Mason)

Meng'en Qin (Mason)

A stats graduate from Henan University, who has a strong interest in 2D vision, diffusion models, multi-modal LLMs.

Patents

Coal and Gangue Intelligent Recognition Based on Selective State Space Equation

Xiaohui Yang, Meng’en Qin, Yanni Zhang, Zheng Ge, Yunfei Bai
Henan University, CN119048740A, 2025

In the invention, we present an intelligent gangue recognition method based on the selective state space modeling, implemented within a lightweight deep learning network. Specifically, the proposed method is composed of a selective state space backbone, SimSPPF module, a multi-scale feature fusion layer, and a decoupled detection head. The state space backbone and SimSPPF module are jointly designed to extract both spatial and semantic information more effectively, thereby enhancing detection precision and robustness. The overall framework enables real-time gangue detection and localization across both still images and video streams, providing comprehensive information such as positions and sizes of coals and gangues. This invention offers strong support for the efficient utilization of coal resources, and is particularly suited for deployment in low-resource or mobile computing environments.

Intelligent Analysis Method of Stomatal Phenotype of Crop Non-destructive Leaves Based on HDIoU

Lichao Peng, Meng’en Qin, Xiaohui Yang, Chen Miao, Yanfeng Sun
Henan University, CN118691972A, 2024

In this invention, we design a triphasic Hellinger Distance based Intersection over Union (HDIoU) for oriented bounding boxes, and apply it to train YOLOv8-OBB on non-destructive leaf stomatal images. HDIoU models each oriented bounding box as a 2D normal distribution and computes the triphasic Hellinger distance between predicted and ground-truth distributions as the box regression loss. It enables models to focus on different objectives at different training periods: in the first phase, the goal is to quickly move the predicted box close to the target; in the second phase, the emphasis shifts to maximizing the overlap between the predicted and ground-truth boxes; in the third phase, the model additionally fine-tunes the predicted box dimensions to closely match the target box’s length and width. The triphasic mechanism does not require manually splitting the training, but leverages a dynamic indicator to automatically determine the current optimization focus during training, allowing the loss function to continuously transition within a single unified learning process. This design enables the model to progressively refine predicted boxes by adjusting its attention—from general proximity to precise alignment of orientation and shape—without interrupting training. Moreover, HDIoU is scale-invariant, making it particularly well-suited for accurately detecting small objects such as stomata.

Multi-modal Data Integration Based on Convolutional Sparse Coding and Optimal Transport

Xiaohui Yang, Jingjing Li, Yuan Feng, Lichao Peng, Yu Song, Meng’en Qin
Henan University and Yellow River Conservancy Technical Institute, CN117763499B, 2024

This invention proposes a multi-modal data integration method based on convolutional sparse coding and optimal transport. We utilize convolutional sparse coding to collaboratively learn the latent representations of two multi-modal features, thereby addressing the interpretability issues commonly associated with neural networks. To better capture the data distribution when generating the latent space, optimal transport theory is employed to mitigate the issue of non-overlapping between the real and generated distributions. Furthermore, alignment is performed based on the correlation matrix between the multi-modal data, and the aligned point pairs are used to fuse the latent spaces of the two modalities, ultimately achieving effective multi-modal data integration. This invention significantly improves the classification accuracy in the integrated latent space and is appropriate for multi-omics integration tasks in the biomedical field, such as single-cell sequencing data and medical imaging data analysis.