| 摘 要: 玉米作为中国关键粮食作物,其品种识别技术随深度学习和计算机视觉技术的发展而进步。针对玉米籽粒品种识别技术已有所发展,但植株品种分类研究较少的不足,提出一种改进的Swin Transformer模型;通过多尺度特征融合和改进的CBAM 注意力机制增强模型的特征表达,提升了模型性能。在自建数据集上的实验结果显示,该方法识别准确率达93.4%,较原模型提高2.7%,且优于 ResNet34、VGGNet、MobileNetV2等模型。以上结果表明,所提方法能够进行高效的玉米植株品种识别。 |
| 关键词: SwinTransformer 多尺度特征融合 CBAM 品种识别 |
|
中图分类号: TP391
文献标识码: A
|
|
| Enhanced Swin Transformer-Based Method for Maize Plant Variety Identification |
|
WANG Shensi, LIU Chengzhong
|
(College Of information Science and Technology, Gansu Agricultural University, Lanzhou 730070, China)
1633630701@qq.com; liucz@gsau.edu.cn
|
| Abstract: As a key grain crop in China, maize variety identification technology has advanced with the development of deep learning and computer vision technologies. Although maize kernel variety recognition techniques
have progressed, research on plant variety classification remains limited. This study proposes an enhanced Swin Transformer approach that strengthens the model’s feature representation through multi-scale feature fusion and an improved CBAM attention mechanism, thereby enhancing model performance. Experimental results on a self-constructed dataset show that the proposed method achieves an identification accuracy of 93.4% , representing a 2.7% improvement over the original model and outperforming models such as ResNet 34, VGGNet, and MobileNet V2. The results demonstrate that the proposed method enables effective maize plant variety identification. |
| Keywords: Swin Transformer multi-scale feature fusion CBAM variety identification |
