Publications

In this work, we identify three critical issues: information loss, context-agnostic sampling and pattern inconsistency in existing multi-scale feature fusion networks and operations. To tackle these issues, this paper proposes Asymptotic Content-Aware Pyramid Attention Network (\(\mathrm{A^3}\)-FPN). Specifically, \(\mathrm{A^3}\)-FPN employs a local-to-global convolutional attention network that gradually enables global feature interaction and disentangles each level from all hierarchical representations. In feature fusion, it collects supplementary content from the adjacent level to generate position-wise offsets and weights for context-aware resampling, and learns multi-scale context reweights to improve intra-category similarity. In feature reassembly, it further strengthens intra-scale discriminative feature learning and reassembles redundant features based on information density and spatial variation of feature maps. Extensive experiments on MS COCO and Cityscapes demonstrate that \(\mathrm{A^3}\)-FPN can easily yield remarkable performance gains on both CNNs and ViTs. Notably, when paired with OneFormer and Swin-L backbone, \(\mathrm{A^3}\)-FPN achieves 49.6 mask AP on MS COCO and 85.6 mIoU on Cityscapes. Furthermore, \(\mathrm{A^3}\)-FPN exhibits powerful capabilities in more precise detection and segmentation, particularly for small，cluttered, and dense objects.

Stomata play a critical role in regulating key physiological processes in plants, particularly for monitoring dynamic changes in living tissues under environmental perturbations. However, accurate high-throughput phenotyping remains bottleneck, as traditional methods rely on destructive sampling and labor-intensive annotations, limiting large-scale and field applications. To solve these challenges, we specially design an oriented detection network for analyzing small, dense, and cluttered stomata in nondestructive images and complex environments. Additionally, a diffusion-based restoration model is introduced to recover blurred images and improve detection accuracy. Restoration and detection are integrated into \(\mathrm{StomaD^2}\), a user-friendly and easy-to-operate all-in-one system that supports nondestructive, in-field stomatal imaging and enables real-time dynamic analysis of stomatal phenotypes (density, conductance, etc.). \(\mathrm{StomaD^2}\) is compatible with both dicotyledonous and monocotyledonous species, destructive and nondestructive images of crop stomata. Experiments show that \(\mathrm{StomaD^2}\) achieves expert-level accuracy, and its generalizability has been validated across diverse plant types, highlighting its potential for large-scale phenotyping, plant physiology research, and precision agriculture.

In this paper, we explore the three-dimensional and topological descriptors of cross-coupling reactions based on the molecular stick-and-ball model and persistent homology analysis, respectively. On this basis, a weighted light CNN with multi-scale subtraction (OS-MSW) is proposed to extract the deep abstract features of the input descriptors, and the extracted features are applied to LightGBM for yield prediction, thus constructing a highly efficient hybrid model, OS-MSWGBM.

This work proposes OCS-TGBM, an intelligent organic chemical synthesis analysis model, which combines topological data analysis (TDA) and Light Gradient Boosting Machine (LightGBM). OCS-TGBM aims at deeply exploring the internal relationship between reaction conditions and yield, and obtaining high-yield reaction conditions and combinations. Additionally, in order to further enhance the generalization of OCS-TGBM, we design a stratified diversity sampling strategy for LightGBM in the training stage. Experiments show that OCS-TGBM is superior to other methods (e.g., XGBoost, SVR and MLPR) in analyzing and predicting the reaction performance of high-throughput organic chemical synthesis.

In neural network models, the perpetual challenge remains in retaining task-relevant information while effectively discarding redundant data during propagation. In this paper, we introduce IB-AdCSCNet, a deep learning model grounded in information bottleneck theory. IB-AdCSCNet seamlessly integrates the information bottleneck trade-off strategy into deep networks, and dynamically adjusts the trade-off hyperparameter λ by FISTA algorithm. By optimizing the compressive excitation loss induced by the information bottleneck principle, IB-AdCSCNet achieves an optimal balance between compression and fitting at a global level, approximating the globally optimal representation feature. This information bottleneck trade-off strategy driven by downstream tasks not only helps to learn effective features of the input data, but also improves the generalization of deep models. Experimental results on CIFAR-10 and CIFAR-100 demonstrate that IB-AdCSCNet not only matches the performance of deep residual convolutional networks but also outperforms them when handling corrupted data. Through the inference of the IB trade-off, the model’s robustness is notably enhanced.