本文为泰国亚洲理工学院(作者：Nampally Tejasri)的硕士论文，共79页。

近几年来，计算机视觉系统一直关注对环境中各种材料的分类和识别，并将其作为一种重要的视觉竞争手段。利用人工神经网络领域的最新发展，了解涉及深度学习过程的不同图像中的材料识别，使我们能够训练各种神经网络结构，以便为这项具有挑战性的任务提取特征。在这项工作中，最先进的卷积神经网络(CNN)技术被用于材料分类，并比较这些不同的结果。收集了两个材料数据集中应用两种流行迁移学习方法的结果。研究结果表明，与完全连接层之前的层所获得的信息有限的情况相比，微调方法取得了很好的效果。比较结果表明，该系统在性能和精度上都有较大的提高，特别是在包含大量图像的数据集中。

The classification and recognition ofvariety of materials that are present in our surroundings become an importantvisual competition have been focused by computer vision systems in the recentyears. Understanding the recognition of the materials in different images thatinvolve a deep learning process made use of the recent development in the fieldof Artificial Neural Networks brought the ability to train various neural networkarchitectures for the extraction of features for this challenging task. In thiswork, state-of-the-art Convolutional Neural Network (CNN) techniques are usedto classify materials and also compare the results obtained by them.The resultsare gathered over two material data sets applying the two popular approaches ofTransfer Learning. The results showcase that fine-tuning approach achieves verygood results compared to the case of approach when the information derived fromthe layer which is just before the fully connected layer is limited. Theresults of the comparison indicates the fact that there is an improvement inthe performance and the accuracy of the system particularly in the data setthat contains large number of images.

1. 引言

2. 文献回顾

3. 研究方法

4. 实验结果

5. 结论

更多精彩文章请关注公众号：

转载本文请联系原作者获取授权，同时请注明本文来自刘春静科学网博客。

收藏

分享

分享到: