A dual-branch multi-modal deep learning framework for non-destructive evaluation of intramuscular fat in sheep

The content of Intramuscular Fat (IMF) is a critical determinant of sheep quality, directly influencing its flavor, tenderness, and juiciness. Although deep learning offers a promising avenue for non-destructive prediction, research has predominantly centered on pork, leaving sheep quality assessment underexplored and highlighting a critical scarcity of public, large-scale multimodal datasets. To overcome the insufficient representational power of single-modality approaches (e.g., B-mode ultrasound images), this paper makes two primary contributions. First, we construct and release a comprehensive multimodal sheep dataset, containing 1,728 samples of ultrasound images, corresponding attributes, and ground-truth IMF values. Second, we propose DB-KAN, a novel dual-branch regression network designed to leverage this rich data. DB-KAN features a Convolutional Neural Network (CNN) branch to extract spatial features from ultrasound images and a Transformer branch to process structured attributes like backfat thickness, eye muscle depth, and eye muscle area measured at the 12th/13th rib site. This dual-branch architecture effectively captures heterogeneous information. Crucially, the decoder innovatively incorporates a KAN-Based regression head (KBRH), which efficiently fuses these multimodal features for a precise final prediction. Experiments on our dataset, partitioned 8:1:1 for training, validation, and testing, demonstrate that DB-KAN achieves state-of-the-art performance. Ablation studies further validate the indispensable roles of both the dual-branch design and the KAN-based fusion strategy.