| 摘 要: 为解决传统神经网络在CIFAR-10(Canadian Institute For Advanced Research)数据集上进行图像分类识别时,存在的模型准确率较低和训练过程易发生过拟合现象等问题,提出了一种将卷积神经网络和批归一化相结合的新神经网络结构构建方法。该方法首先对数据集进行数据增强和边界填充处理,其次对典型的 CNN(Convolutonal Neural Networks)网络结构进行改进,移除了卷积层组中的池化层,仅保留了卷积层和BN(Batch Normalization)层,并适量增加卷积层组。为了验证模型的有效性和准确性,设计了6组不同的神经网络结构对模型进行训练。实验结果表明,在相同训练周期数下,推荐使用的 model-6模型表现最佳,测试准确率高达90.17%,突破了长期以来经典CNN在CIFAR-10数据集上难于达到90%准确率的瓶颈,为图像分类识别提供了新的解决方案和模型参考。 |
| 关键词: 图像分类识别;卷积神经网络;批归一化;数据增强;边界填充 |
|
中图分类号: TP391.4
文献标识码: A
|
| 基金项目: 广东省普通高校重点领域专项“基于GAN人脸超分辨率恢复关键技术的研究”(2021ZDZX1101);广东省教育科学规划课题高等教育科学研究专题“基于‘岗课赛证’融通的专业群课程体系建构与特色发展研究”(2023GXJK948) |
|
| Research on Image Classification and Recognition Method Based on BatchNormalization Convolutional Neural Network Algorithm |
|
XIE Zhiming1,2
, GU Fang1,2
|
(1.School of Engineering, Shanwei Institute of Technology, Shanwei 516600, China;
2.Shanwei Innovation Industrial Design and Research Institute, Shanwei 516600, China)
106710856@qq.com; 505665578@qq.com
|
| Abstract: To address the issues of low model accuracy and frequent overfitting during training when traditional neural networks perform image classification and recognition on the CIFAR-10 (Canadian Institute For Advanced Research) dataset, this paper proposes a novel neural network architecture construction method combining Convolutional Neural Networks (CNNs) with Batch Normalization (BN). The method first applies data augmentation and boundary padding to preprocess the dataset, then modifies the conventional CNN architecture by removing pooling layers from convolutional layer groups while retaining convolutional layers and BN layers, with appropriate increases in convolutional layer depth. To validate the effectiveness and accuracy of the model, six distinct neural network architectures are designed for comparative training experiments. Experimental results demonstrate that under identical training epochs, the recommended Mode-l 6 achieves optimal performance with a test accuracy of 90.17% , breaking through the long-standing bottleneck where classical CNNs struggled to exceed 90% accuracy on the CIFAR-10 dataset. This provides new solutions and model references for image classification and recognition tasks. |
| Keywords: image classification and recognition; Convolutional Neural Network (CNN); Batch Normalization
(BN); data augmentation; boundary padding |
