| 罗富贵,李明珍.基于卷积核分解的深度CNN模型结构优化及其在小图像识别中的应用[J].井冈山大学自然版,2018,(2):31-39 |
| 基于卷积核分解的深度CNN模型结构优化及其在小图像识别中的应用 |
| THE OPTIMIZATION OF SUPER DEEP CNN MODEL BASED ON CONVOLUTIONAL KERNEL DECOMPOSITION AND APPLICATION ON TINY IMAGES RECOGNITION |
| 投稿时间：2017-12-30 修订日期：2018-02-28 |
| DOI：10.3969/j.issn.1674-8085.2018.02.006 |
| 中文关键词: 卷积神经网络 卷积核分解 小图像 识别 超深度模型 |
| 英文关键词: convolutional neural networks convolutional kernel decomposition tiny image recognition super deep model |
| 基金项目:河池学院2016年校级重点项目（XJ2016ZD007）；河池学院2017年校级重点项目（XJ2017ZD08） |
|
| 摘要点击次数: 1573 |
| 全文下载次数: 2409 |
| 中文摘要: |
| 小图像由于像素少、分辨率低、整幅图像包含信息较少，识别较为困难。目前优秀的深度卷积神经网络模型多为大图像而设计，而用于小图像的模型则存在着层次不够深、难以对特征进行充分抽象的不足。本文基于VGG19模型，依据卷积核分解的原理，设计了一种KDS-DCNN模型，模型深度达到31层，解决了目前超深度模型不能直接用于小图像识别的问题，实验表明该方法不但提升了识别性能，而且还降低了模型的时间复杂度。在CIFAR-10、CIFAR-100和SVHN三个数据集上的验证结果显示，KDS-DCNN模型性能优越，其识别错误率分别降低到29.46%、6.02%和2.17%。 |
| 英文摘要: |
| The tiny image recognition is one of important parts of computer vision. But it is very difficult to recognize them because of their few pixels, low resolution and poor information. The deep convolutional neural networks have obtained great breakthrough on many vision tasks such as object classification and detection via more than one times of linear and non-linear transformation. However, the popular outstanding deep CNN models are always for large images, and the models for tiny images always have shallow depth and insufficient abstract features. In this work, a novel super deep CNN model named KDS-DCNN for tiny images is proposed according to the principle of convolutional kernel decomposition based on the popular VGG19 model, and the model depth reaches to 31 and is with low complexity. The experimental results demonstrate that the model has better performance than other same kinds of models. And the state-of-the-art results are obtained on CIFAR-10, CIFAR-100 datasets, and the recognition error rate drops down to 29.64% and 6.02% respectively. And on the SVHN dataset, a competitive result is obtained too, the recognition error rate drops down to 2.17%. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
