To enhance the precision and reliability of Optic Cup and Optic Disc Segmentation, we present a novel Multi-scale Channel Attention Vision LSTM (MCA-ViLSTM) Network. It integrates the encoder-decoder framework of U-Net with the sequential modeling capabilities of Long Short-Term Memory (LSTM) networks, thereby enabling the model to simultaneously capture local morphological features and long-range spatial dependencies while maintaining computational efficiency. Furthermore, we introduce a Multi-scale Channel Attention (MCA) module that enhances feature representation for small anatomical structures through adaptive channel weighting mechanisms, effectively addressing segmentation challenges arising from class imbalance. Comprehensive experiments conducted on publicly available datasets (DRIONS-DB and REFUGE) demonstrate that MCA-ViLSTM achieves superior segmentation performance in fundus image analysis.
Recent text-conditioned image generation models have demonstrated an exceptional capacity to produce diverse and creative imagery with high visual quality. However, when trained on billion-scale datasets randomly collected from the Internet, where human aesthetic preferences are inadequately learned, these models often generate facial images that deviate from mainstream aesthetics, particularly across different racial groups. While some existing methods attempt to address this issue by fine-tuning models on large-scale, manually annotated facial datasets, such approaches incur substantial annotation and computational costs. To overcome these limitations, we propose a framework called FP-Director, which learns a facial-preference direction in the latent space and updates latent codes to refine the generated results. This alignment process is applied during inference and does not require fine-tuning of the original model or access to large-scale datasets. Extensive empirical evaluations demonstrate that FP-Director significantly improves both overall quality and aesthetics of generated faces.
Cattle and sheep farming play a pivotal role in agricultural economies, yet parasitic diseases significantly impede industry development. Traditional manual diagnostic methods suffer from low efficiency and subjectivity. While art ificial intelligence (AI) based detection shows promise, its advancement is constrained by limited data availability. This study aims to integrate microscopic images, morphological characteristics and genomic data of bovine and sheep parasites, and establish a multimodal biometric database of bovine and sheep parasites. Standardized protocols were implemented to collect blood, fecal, and ectoparasite samples. Expert teams conducted cross-validated annotations following unified protocols to ensure data integrity. The database system, developed based on MySQL, comprises a structural design, a user front-end module, and an administrator backend module. This database provides comprehensive data support for AI model development and validation, promising enhanced efficiency and accuracy in parasitic disease detection, which facilitates the healthy development of the animal husbandry industry.
In clinical medicine, the accuracy of sleep staging directly affects the diagnosis and treatment of sleep disorders. However, existing methods have significant limitations in characterizing dynamic spatiotemporal associations of the brain during sleep, especially in modeling nonlinear interactions across brain regions. To overcome these limitations, we develop STHGNN, a hypergraph-based neural network model that captures spatiotemporal features for automated sleep staging. The model fuses temporal and spatial features extracted from multi-modal data, and thus performs classification task effectively. In our experiments, we compared our method on the ISRUC-S1 and ISRUC-S3 datasets with the latest methods. Experimental results on the ISRUC-S3 dataset show that the overall classification accuracy of this method reaches 82.8
Addressing challenges in UAV-based pedestrian detection, including small target sizes, complex backgrounds, and computational constraints of resource-limited devices, this study proposes an improved pedestrian detection algorithm, CGDRF-YOLO. First, the algorithm proposes a downsampling module to aggregate multi-scale receptive field and context information, which mitigates feature loss and strengthens anti-interference capabilities. Secondly, a content-aware feature reassembling operator is integrated to dynamically generate adaptive kernels and reorganize features, thus enhancing feature utilization efficiency. Finally, the Inner-WIoU loss function is incorporated to improve localization precision and optimize loss calculation. Experimental results from the VisDrone2019 pedestrian detection sub-dataset demonstrate that, compared with the baseline algorithm, CGDRF-YOLO not only improves mAP50, mAP50:95 and FPS by 3.6
