| 摘 要: 针对现有马铃薯块茎品种识别任务中,数据集规模较小、现有模型分类准确率和推理效率较低等问题,提出一种基于改进ResNet50的分层混合注意力网络模型ECBAM-ResNet50。该模型以ResNet50为基础,在Bottleneck内部嵌入ECA模块进行局部通道优化,并在layer2及layer3后引入轻量化CBAM模块进行全局空间-通道特征提取。实验基于97个品种、43645张图像的大规模马铃薯块茎数据集。实验结果表明,该模型在测试集上的分类准确率达到99.16%,精确率、召回率、F1值也表现优异,同时相较于MobileNetV3、DenseNet121及Vision Transformer等模型准确率有显著提高。该研究表明,改进后的模型能够进行高效精准的马铃薯块茎品种识别,同时也为深度学习在作物品种识别领域的改进与应用提供了新思路。 |
| 关键词: 马铃薯块茎 Resnet50 品种识别 注意力机制 |
|
中图分类号:
文献标识码:
|
| 基金项目: 甘肃省高等学校产业支撑计划项目 |
|
| Potato Tuber Variety Recognition Based on an Improved ResNet50 Model |
|
yang ya wen, liu cheng zhong, zhang lin zhe
|
College of Information Science and Technology, Gansu Agricultural University
|
| Abstract: Addressing challenges in potato tuber variety identification—such as limited coverage of actual cultivated varieties in small-scale datasets and inadequate feature discrimination capabilities of existing models for multi-variety tubers, which result in low classification accuracy and inference efficiency—this study proposes an enhanced ResNet50-based hierarchical hybrid attention network model, ECBAM-ResNet50. Building upon ResNet50, this model embeds an ECA module within the Bottleneck layer for local channel optimization. Lightweight CBAM modules are introduced after layers 2 and 3 to perform global spatial-channel feature extraction, enabling efficient and precise potato tuber variety classification. Experiments were conducted using a large-scale potato tuber dataset comprising 97 varieties and 43,645 images. Experimental results demonstrate that the ECBAM-ResNet50 model achieves a classification accuracy of 99.16% on the test set, with outstanding precision, recall, and F1 scores. It also shows significant improvements in accuracy compared to mainstream models such as MobileNetV3, DenseNet121, and Vision Transformer. This research demonstrates that the improved model enables efficient and precise multi-variety potato tuber classification, while also offering new insights for enhancing and applying deep neural networks in crop variety identification. |
| Keywords: potato tubers ResNet50 variety identification attention mechanism |
