【摘要】 针对家具流通环节木材识别效率低、端侧算力受限的问题，文章提出了兼顾精度与轻量化的BEL-ShuffleNet网络模型，为木材贸易、质量抽检以及消费者判断家具材质提供了一种便捷的方法。文章以ShuffleNet V2为基础网络，实施了3项定向优化：采用模糊池化BlurPool替代最大池化抑制混叠；在Squeeze-and-Excitation框架下设计高效归一化注意力模块(Efficient Normalization-based Attention Module, ENAM),取代多层感知机进行通道重标定；设计轻量化特征融合模块(Lightweight Channel-Lifting Module, LCLM)以替代末端卷积，并通过多组对照与消融实验验证其有效性。在测试集上，BEL-ShuffleNet准确率达94.95%,较基线提升2.88%。相比经典卷积神经网络MobileNetV4(88.94%)、EfficientNet-Lite(92.71%)、ResNet 18(91.51%)、RepVGG-A0(90.79%)和GoogLeNet(91.83%),各精度指标均为最优，推理时延也与ShuffleNet V2、MobileNetV4属同一量级。所提BEL-ShuffleNet模型，在保持轻量化的同时，提升了木材识别精度，为家具木材品类在移动端实现快速、无损识别提供了可行基础。更多还原

【Abstract】 To address the low efficiency of wood identification in furniture circulation and the limited computing power of edge devices, a lightweight BEL-ShuffleNet model was proposed based on ShuffleNet V2. 3 improvements were introduced: BlurPool was used to suppress aliasing during downsampling, the efficient normalization attention module(ENAM) was designed for channel recalibration, and the lightweight channel-lifting module(LCLM) was adopted for feature fusion. On the self-constructed test set, BEL-ShuffleNet achieved an accuracy of 94.95%, 2.88% higher than the baseline. Compared with MobileNetV4, EfficientNet-Lite, ResNet18, RepVGG-A0, and GoogLeNet, the proposed model obtained the best overall recognition performance while maintaining competitive inference efficiency. The model provides a feasible basis for rapid and non-destructive mobile-side recognition of furniture wood species.更多还原