基於深度摺積神經網路之影像檢索技術

Using Deep Convolutional Neural Networks for Image Retrieval

周揚賀(Yang-Ho Chou)

指導教授 ： 張寶基

國立中央大學/資訊電機學院/通訊工程學系/碩士(2015年)

摘要

隨著大數據(big data)時代的來臨，人類產生的數位化資料量每天以爆炸式地增長。在面對鉅量影音資料衝擊，如何有效管理日益龐大的多媒體資料庫，並根據使用者的需求取得所需之影像，是目前影像檢索領域所面臨的重要課題。影像檢索技術發展歷史悠久，包含基於文字或內容之影像檢索皆有廣泛的研究與應用，卻也面臨一些使用上的問題。基於文字之影像檢索需要人工註解而耗時耗力，且容易受到標記文字描述的主觀程度所影響；而基於內容之影像檢索則因為傳統影像內容特徵描述無法完整表達影像人類所感知的高階語意概念，使得檢索效果往往不如使用者的預期。
本論文利用深度學習中的摺積神經網路(convolutional neural networks, CNN)做為影像特徵學習的方法。能藉由許多不同層級的特徵處理，將影像內容轉換成抽象的深層語意概念，更能夠代表影像的關鍵特徵，讓使用者能迅速取得正確的檢索影像結果，有助於改善現有影像搜索技術之發展。由實驗結果顯示，CNN的深層架構確實能夠逐層學習出影像的關鍵特徵描述，影像分類的準確度高於一般神經網路。而將所學習的特徵描述應用到影像檢索中，CIFAR-10影像資料庫的檢索平均準確率(mean average precision, MAP )達到0.707的不錯表現。

關鍵字

影像檢索 ； 深度學習 ； 特徵學習 ； 類神經網路 ； 摺積神經網路

並列摘要

In the age of Big Data, with the rapid development of Internet and mobile devices, people can easily obtain audio and video information everywhere. Because of the exponential growth of multimedia data, how to efficiently retrieve and manage multimedia information from huge databases becomes an important issue on image retrieval. In content-based image retrieval (CBIR), most existing hand-crafted feature descriptors were considered low-level and far from what human normally perceived the world. It is a challenging problem of “semantic gap” between low-level visual features captured by machines and high-level semantic concepts perceived by human.
This thesis focuses on the high-level image feature learning by the convolutional neural networks (CNN). CNN, as a deep learning framework, can extract image features in different layers, and transfer the image content into (abstract) semantic concepts. These high-level features descriptors can be better image representations than the hand-crafted feature descriptors, and further improve the image retrieval performance. The experimental results show that layer-wise learning of feature hierarchies in CNN can represent the images very well. Using CNN for feature extractions performed on CIFAR-10 dataset with 79.3% accuracy in image classification task, and with 0.707 of mean average precision (MAP) in image retrieval tasks.

並列關鍵字

Content-based image retrieval ； Deep learning ； Feature learning ； Neural networks ； Convolutional neural networks

參考文獻

[2] G. Casella, E. I. George, “Explaining the Gibbs Sampler,” The American Statistician, vol.46, no. 3, pp. 167-174, 1992.

[6] G. E. Hinton, et al., “Deep Neural Networks for Acoustic Modeling in Speech Recognition,” IEEE Signal Processing Magazine, November, 2012.

[7] C. Neubauer, “Evaluation of convolutional neural networks for visual recognition,” IEEE Transactions on Neural Networks, vol. 9, no. 4, pp. 685-696, July 1998

[9] Y. Lecun, L. Bottou, Y. Bengio and P. Haffner, “Gradient-based learning applied to document recognition”, Proceedings of the IEEE, vol. 86, no. 11, pp. 2278-2324, 1998

[10] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks”, Advances in neural information processing systems, pp. 1097-1105, 2012.