<p>Effective identification of unsound wheat kernels is crucial for ensuring grain quality and optimizing storage management. Existing methods based on Convolutional Neural Networks (CNNs) have limited ability to capture global contextual dependencies. Transformers can model long-range dependencies through self-attention mechanisms, but they often lack sensitivity to fine-grained local features. The reliability of identifying unsound wheat kernels depends on the effective capture and fusion of both global and local features. To address this issue, this paper proposes a dual-branch feature fusion network named ETFNet, which integrates local details and global semantics. The model extracts fine-grained texture features via an EfficientNetV2 branch while capturing long-range dependencies via a Swin Transformer branch. A Multi-scale Feature Fusion Module (MFFM) is designed to hierarchically fuse features from the two branches. Additionally, a Feature Enhancement Module (FEM) integrates attention mechanisms with inverted residual structures to suppress noise and enhance discriminative feature representation. To validate the effectiveness of ETFNet, experiments were conducted on the public GrainSpace dataset, which includes normal kernels (NOR) and six types of unsound kernels: Fusarium and Shrivelled (FS), Sprouted (SD), Moldy (MY), Broken (BN), Attacked by Pests (AP), and Black Pointed (BP). Experimental results demonstrate that ETFNet achieves high accuracy and outperforms several mainstream methods. This research provides an efficient and reliable solution for the automated detection of unsound wheat kernels, offering significant potential to reduce storage losses, ensure food security, and advance smart agriculture applications.</p>