植基於適應性特徵之影像查詢系統與影像分類研究

Study of Image Retrieval System and Image Classification Based on Adaptive Features

吳育陞(Yu-Shung Wu)

指導教授 ： 林春宏

國立臺中科技大學/資訊學群/資訊科技與應用研究所/碩士(2010年)

https://doi.org/10.6826/NUTC.2010.00017

摘要

本研究主要針對影像提出顏色(color)和紋理(texture)特徵的萃取(extraction)，並分別應用於影像查詢(image retrieval)系統與影像分類(image classification)上的研究兩種。
本文提出的特徵萃取有改良K-means顏色分群的方法，以及適應性共生矩陣之圖形(Adaptive Motifs Co-Occurrence Matrix, AMCOM)、梯度適應性圖形之直方圖(Gradient Histogram for Adaptive Motifs, GHAM)兩種萃取紋理特徵的方法。文中改良K-mean顏色分群的方法分別以刪除未有像素(pixel)歸屬之群組、相似群中心的合併以及建立偏離像素之群組等三種處理步驟來改善K-means演算法。AMCOM用於計算影像中相鄰像素之間可能產生的像素差異關係，再統計此像素差異關係，做為描述影像紋理的特性。GHAM是統計樣本區塊圖形(motif of pattern block)內的梯度(gradient)平均值，主要是描述每一個樣本區塊圖形内的紋理情況。
影像查詢部份則先行探討不同的特徵挑選方法，其中包含前向式特徵挑選(Sequential Forward Selection, SFS)、後向式特徵挑選(Sequential Backward Selection, SBS)以及本文所提出修改後的基因演算法之特徵挑選(Genetic Algorithms Feature Selection, GAFS)做相互比較及分析。之後將所挑選的結果進行影像查詢，並與其它學者所提的特徵萃取方法應用於影像查詢相互比較其實驗結果；影像分類則針對於本文所萃取獲得之影像特徵，透過目前學術界較常使用的分類法(classification)--向量支援機(Support Vector Machine, SVM)，來測試本文所使用的影像特徵與其它學者的影像特徵做分析與比較，並探討其實驗結果。
在最後實驗結果可以發現，本文所提及針對於影像的特徵萃取，都能有助於影像查詢及影像分類的結果。而在透過基因演算法之特徵挑選運用於影像查詢，除了特徵個數能有效減少之外，其影像查詢結果的成功率仍然優於其它學者所提的方法，能證明特徵數不在於數量愈多愈好，而是能把有助於影像描述的特徵給予保留，其影像查詢結果仍然可以達到不錯的水準。

關鍵字

K-mean演算法 ； 紋理特徵 ； 基因演算法 ； 特徵挑選 ； 向量支援機

並列摘要

In this study, the application of image retrieval system and image classification is the use of it in color and texture feature extraction.
This research proposed by feature extraction has improved K-means and genetic algorithm method of the two color cluster, and AMCOM (Adaptive Motifs Co-Occurrence Matrix), GHAM (Gradient Histogram for Adaptive Motifs), Extract texture characteristics of two methods. Inside the article improve K-mean color cluster approach, respectively, to remove not been assigned to the pixels of the group, similar group center of the merger and the establishment to deviate from the pixels of the group, three processing steps to improve the K-means algorithm. Genetic algorithms applied to color image grouping method of, The group number and the initial group center set to chromosome, re-use of K-means algorithm to do the pixel values of grouping, the end, fitness function as a group number of the assessed value, AMCOM used to calculate the image That may arise differences in the relationship between pixel, Then statistical difference between the pixel, As described image texture features. GHAM statistical motif of pattern block gradient of the average, mainly describe a situation of each motif of pattern block within the texture.
As to the image retrieval part, different feature selection methods would be investigated first, including Sequential Forward Selection (SFS), Sequential Backward Selection (SBS), and modified Genetic Algorithms Feature Selection (GAFS) proposed in this study, and then compared as well as analyzed with each other. After that, the chosen results would be used for image retrieval, and the experiment results would be compared with feature extraction method proposed by other scholars which is applied on image retrieval; For image classification, image features extracted in this study would be testified by the classification usually used in current academic -- Support Vector Machine (SVM), and analyzed/compared with image features extracted by other scholars to investigate the experiment results.
In the final experiment results, we can find that all the methods about feature extraction mentioned in this study can contribute to the better results of image retrieval and image classification. By applying GAFS on image retrieval, not only the number of features could be reduced effectively, but also a higher image-retrieval hit rate than that of other scholars’ methods could be obtained. This can prove that the number of features is not the more the better. If the useful features for describing images could be reserved, good image retrieval results are still achievable.

並列關鍵字

K-mean algorithm ； texture features ； genetic algorithms ； feature selection ； support vector machine

參考文獻

[2] Y. Gu, B. Panda, and K. A. Haque, “Design and Analysis of Data Structures for Querying Image Databases,” ACM Symposium on Applied Computing, Las Vegas, Nevada, USA, pp. 236-241, March 2001.

[3] V. Vapnik, “Structure of Statistical Learning Theory,” Computational Learning and Probabilistic Reasoning, 1996.

[6] Y. Rui and T. S. Huang, “Image Retrieval: Current Techniques, Promising Directions, and Open Issues,” Journal of Visual Communication and Image Representation, Vol. 10, pp. 39-62, October 1999.

[7] C. C. Chang and Y. K. Chan, “A Fast Filter for Image Retrieval Based on Color Features,” SEMS2000, Baden-Baden, German, pp. 47-51, July 2000.

[8] R. Brnuelli and O. Mich, “Histograms Analysis for Image Retrieval,” Pattern Recognition, Vol. 34, No. 8, pp. 1625-1637, Auguest 2001.