In modern animal husbandry, the eye muscle area is a key meat quality indicator, influencing consumer choices. Traditionally, measuring this area involves post-slaughter manual tracing, which is time-consuming and labor-intensive. In this work, due to a lack of public animal ultrasound datasets, this study established the EM-Sheep dataset for automated eye muscle in sheep using ultrasound images. We propose an advanced ultrasound image segmentation algorithm, MOC-YOLO, based on multi-scale feature fusion and attention mechanisms. This algorithm builds upon YOLOv8 as the base network and integrates the MOCA module to enhance multi-scale feature extraction and address complex boundary recognition in the segmentation of the eye muscle area. Additionally, the C2f-CGA module addresses noise, low contrast, and artifacts in B-mode ultrasound images. Experimental results show that MOC-YOLO significantly improves segmentation performance on the EM-Sheep dataset, with a 2.4% increase in IoU, 7.2% in accuracy, 3.9% in Dice and F1 scores, and boosts in Precision and Recall by 2.4% and 5.3%, respectively.

In medical image segmentation tasks, convolutional neural networks (CNNs) have been extensively employed due to their exceptional ability to extract local features. In recent years, Vision Transformers (ViTs) have also gained prominence in medical image segmentation, leveraging their powerful capacity to model long-range dependencies. However, CNNs encounter significant limitations in capturing global dependency information, while the quadratic growth in computational complexity with sequence length in ViTs hampers their efficiency in practical applications.To address this bottleneck, state space models (SSMs), such as Mamba, have recently demonstrated remarkable performance in modeling long-range dependencies with linear computational complexity. Despite this advantage, Mamba faces challenges in effectively capturing fine-grained local features, which are essential for high-precision segmentation tasks. To overcome these challenges, this paper introduces a novel dual-stream U-shaped network architecture, termed VMC-UNet. The proposed model integrates the efficient global modeling capabilities of Mamba with the precise local feature extraction strengths of CNNs, offering a balanced and effective solution.Experimental results reveal that VMC-UNet achieves significant improvements in segmentation accuracy and robustness across various medical image segmentation tasks, including organ and skin lesion segmentation. .

With the development of various online platforms, research into recommender systems has advanced significantly. However, recommender systems still face significant issues of data sparsity and noise. To address these challenges, Graph Contrastive Learning (GCL) has emerged, leveraging data augmentation and Graph Neural Networks (GNN) to enhance Recommender performance. However, existing GCL methods, which rely on randomly disrupting the graph structure or manually designed contrastive tasks, can easily lead to information loss and difficulties in capturing complex user-item interactions. To address the limitations above, this paper proposes a novel model architecture called DiffGFormer, which integrates the strengths of graph neural networks and Gaussian Diffusion Models. DiffGFormer aims to learn generative representations of user preferences. It can also accurately capture complex higher-order relationships and combines collaborative filtering with graph structure information to enhance the ability to capture interaction patterns and generate new Recommenders. Additionally, this paper introduces a novel data augmentation module—graph augmentation and regularization strategy module (EDG)—which adjusts edge weights based on node degree to avoid oversimplifying the graph structure. Experimental results validate the superiority of the DiffGFormer model in Recommender tasks.

In modern animal husbandry, artificial intelligence helps accurately manage individual sheep. However, it is difficult to recognize the sheep’s facial features and capture the nuances. It is not easy to extract the fine-grained features of a sheep’s face because the traditional vision transformer cannot realize the effective embedding of the interaction scale. To address this problem, we propose a novel sheep Transformer tool called SheepFormer . This model comprises components such as the Interactive Scale Embedded Images Block ( ISEI ), Patch Short Long Distance Attention Module ( PSLDA ), Dynamic Relative Position Offset ( DRPO ), and Transformer Neck and Head ( TNH ). These components are designed to embed features at multiple scales, fuse long and short-distance self-attention, adaptively handle relative position offsets for various group sizes, and introduce a prediction head to detect fine-grained facial targets in sheep at different scales. SheepFormer integrates Residual Attention to seek dense facial features in sheep and utilizes a Transformer Head to replace the traditional head, exploring the predictive potential of self-attention mechanisms in sheep faces. Experimental results demonstrate a 7.1% improvement in average precision (AP) for sheep face detection compared to Collaborative DEtection TRansformer (CO-DETR) and an 11.12% enhancement in accurate classification of sheep identity document (ID) compared to Shifted Windows Transformer (Swin Transformer). This study demonstrates that SheepFormer can extract fine-grained facial features of sheep, which promotes the advancement of high-precision sheep individual recognition and provides a guide for recognizing kinship.