| 摘 要: 针对传统小麦籽粒检测任务中种类少、数目小、识别效率不高、受数据集因素影响较大的问题,构建30000张小麦籽粒图像数据集进行分类研究。在原始 ResNet34模型的残差结构中加入改进的SE-P注意力机制,减少无关的特征依赖,增强模型的特征表达能力;在全连接层之前应用 Dropout层,通过随机丢弃部分神经元,降低过拟合的发生。实验结果表明,改进后的 ResNet34分类模型准确率、精确度和召回率分别为92.30%、92.23%和92.72%,相较于原模型准确率提升3.71%。在小麦籽粒分类任务中提升明显。 |
| 关键词: 小麦籽粒 ResNet34 Dropout层 SE-P注意力机制 |
|
中图分类号: TP391.7
文献标识码: A
|
| 基金项目: 国家自然科学基金项目(32360437);甘肃省高等学校创新基金项目(2021A-056) |
|
| Research on Wheat Kernel Classification Based on the Improved ResNet34 Network Model |
|
XUE Hao1, LIU Chengzhong1, LU Qinglin2
|
(1.College of Information Science and Technology, Gansu Agricultural University, Lanzhou 730070, China;
2.Wheat Research Institute, Gansu Academy of Agricultural Sciences, Lanzhou 730070, China)
296398279@qq.com; liucz@gsau.edu.cn; gsnklql@gsagr.ac.cn
|
| Abstract: To address the limitations of traditional wheat kernel detection methods—such as limited variety coverage, smal-l scale datasets, low recognition efficiency, and high susceptibility to dataset biases—this study constructs a dataset of 30 000 wheat kernel images for classification tasks. An enhanced SE-P attention mechanism was incorporated into the residual blocks of the original ResNet34 model to reduce irrelevant feature dependencies and strengthen feature representation capability. Additionally, a Dropout layer was applied before the fully connected layer to mitigate overfitting by randomly deactivating neurons. Experimental results demonstrate that the improved ResNet34 model achieves accuracy, precision, and recall rates of 92.30% , 92.23% , and 92.72% , respectively, representing a
3.71% increase in accuracy compared to the baseline model. Significant improvements were attained for wheat kernel classification tasks. |
| Keywords: wheat kernels ResNet34 Dropout layer SE-P attention mechanism |
