基於旋積類神經網路的影像搜尋應用中設計具鑑別能力的影像區塊挑選法

Discriminatively-learned Patch Selection for Image Retrieval with CNN

古韋麟(Wei-Lin Ku)

指導教授 ： 彭文孝

國立交通大學/資訊學院/多媒體工程研究所/碩士(2015年)

https://doi.org/10.6842/NCTU.2015.00473

摘要

為了提升的影像搜尋的精確度，本論文提出一個具鑑別能力的影像區塊挑選
法，以利後續旋積類神經網路(CNN)的特徵萃取並形成一個更強化的全域影像表
示法。本論文的全域影像表示法架構以MOP-CNN 為基礎，我們把先它的滑動
視窗影像區塊偵測換成以物件為導向的偵測法，以取得更多有用並且具鑑別力的
資訊。接著，為了要捨棄多餘的影像區塊(帶有過多雜訊)，我們設計了一種針對
影像搜尋具鑑別力的影像區塊挑選法。經過篩選後的影像區塊被輸入至旋積類神
經網路以萃取影像特徵。而這些多個特徵可以通過一個簡單的最大匯集
(Max-pooling)運算以形成單一個全域影像表示法。實驗結果顯示經過所設計影像
區塊挑選法所形成的全域影像表示法在四個著名的影像搜尋資料庫贏過
MOP-CNN以及以CNN為全域特徵擷取的表示法。

關鍵字

影像檢索 ； 視覺搜尋 ； 旋積類神經網路 ； 物件辨識

並列摘要

In this thesis, targeting at advancing image retrieval tasks, we propose a
discriminatively-learned patch selection process that improves the global image
representation with local-CNN-features. Starting from MOP-CNN, we replace its
sliding-window mechanism with the objectness estimator for producing local image
patches with more useful and discriminative information. Then, to further discard
noisy patches, we propose to incorporate a discriminative patch selection mechanism.
Strictly selected patches are passed through the CNN to extract discriminative
features, from which the simpler MAX-pooling can be performed on those features.
The result is a global image representation with more discriminative power and even
shorter feature dimension. Experimental results show that when compared with
MOP-CNN and CNN as a global feature extractor, our proposed framework brings
significant performance benefits on four retrieval datasets.

並列關鍵字

Image Retrieval ； Visual Search ； Convolutional Neural Network ； Object Recognition

參考文獻

[5] P. Arbeláez, J. Pont-Tuset, J. Barron, F. Marques, and J. Malik, “Multiscale combinatorial grouping,”in CVPR, 2014.

[6] A. Babenko, A. Slesarev, A. Chigorin, and V. S. Lempitsky, “Neural codes for image retrieval,”in ECCV, 2014.

[7] M.-M. Cheng, Z. Zhang, W.-Y. Lin, and P. H. S. Torr, “BING: Binarized normed gradients for objectness estimation at 300fps,”in CVPR, 2014.

[8] P. Dollár and C. L. Zitnick, “Structured forests for fast edge detection,”in ICCV, 2013.

[9] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman, “The PASCAL Visual Object Classes Challenge 2007 (VOC2007) Results,”http: //www.pascal-network.org/challenges/VOC/voc2007/workshop/index.html.