基于空洞卷积残差网络的作物害虫图像分类方法

Image Classification Method for Crop Pests Based on Hollow Convolutional Residual Networks

摘要

针对传统卷积神经网络在作物害虫图像分类中易过拟合、泛化能力不足的问题，提出一种基于空洞卷积残差网络（Dilated Convolution Residual Network, DCResNet）的分类方法。系统评估ResNet-18、ResNet-34和ResNet-50在害虫图像数据集上的性能，分析模型深度、超参数及数据增强对分类效果的影响。实验结果表明，ResNet-18达到82.16%的识别准确率，优于ResNet-34（79.83%）和ResNet-50（71.24%）。最优超参数组合为学习率0.001、批量大小64，数据增强使准确率提升6.23%。该方法可为农作物害虫智能识别提供有效技术支持。

Abstract

To address the issues of overfitting and insufficient generalization capability inherent in traditional convolutional neural networks for crop pest image classification, this study proposes a classification method based on the Dilated Convolution Residual Network (DCResNet). The performance of ResNet-18, ResNet-34, and ResNet-50 was systematically evaluated on a pest image dataset, with analyses conducted on the impact of model depth, hyperparameters, and data augmentation on classification accuracy. Experimental results demonstrate that ResNet-18 achieves an identification accuracy rate of 82.16%, surpassing ResNet-34 (79.83%) and ResNet-50 (71.24%). The optimal hyperparameter configuration is learning rate 0.001 and batch size 64; data augmentation further improves accuracy by 6.23%. This approach provides robust technical support for intelligent crop pest recognition.

关键词

残差网络 / 害虫分类 / 图像识别

Key words

residual network / pest classification / image recognition

参考文献

[1] 孙俊, 谭文军, 毛罕平, 等. 基于改进卷积神经网络的农作物病虫害识别研究[J]. 农业机械学报, 2017, 48(11): 78-85.

[2] 张建华, 孔繁涛, 吴建寨, 等. 基于改进VGG16网络的农作物害虫识别方法[J]. 农业工程学报, 2019, 35(10): 148-155.

[3] 周飞燕, 金林鹏, 董军. 卷积神经网络研究综述[J]. 计算机学报, 2017, 40(6): 1229-1251.

[4] Szegedy C, Ioffe S, Vanhoucke V, et al. Inception-v4, inception-resnet and the impact of residual connections on learning[C]. AAAI Conference on Artificial Intelligence. 2017, 31(1).

[5] 蒋心璐, 陈天恩, 王聪, 等. 农业害虫检测的深度学习算法综述[J]. 计算机工程与应用, 2023, 59(6): 30-44.