A robust local sparse coding method for image classification with Histogram Intersection Kernel

详细信息    查看全文

• 作者：Pan LiAuthor Vitae ; Yang Liu ; ^{yliu76@hit.edu.cn" class="auth_mail" title="E-mail the corresponding author}Author Vitae ; Guojun LiuAuthor Vitae ; Maozu Guo ^{maozuguo@hit.edu.cn" class="auth_mail" title="E-mail the corresponding author}Author Vitae ; Zhiyong PanAuthor Vitae
• 关键词：Local sparse coding ; Histogram Intersection Kernel ; Image classification ; Non-negative sparse coding
• 刊名：Neurocomputing
• 出版年：2016
• 出版时间：5 April 2016
• 年：2016
• 卷：184
• 期：Complete
• 页码：36-42
• 全文大小：1215 K

文摘

Local sparse coding methods have been shown to lead to increased performance in image classification when it takes histograms as inputs. These methods often use Euclidean (pan id="mmlsi0003" class="mathmlsrc">pan class="formulatext stixSupport mathImg" data-mathURL="/science?_ob=MathURL&_method=retrieve&_eid=1-s2.0-S0925231215018962&_mathId=si0003.gif&_user=111111111&_pii=S0925231215018962&_rdoc=1&_issn=09252312&md5=3c06b0195e62db7e753444a440b32cc7" title="Click to view the MathML source">l₂pan>pan class="mathContainer hidden">pan class="mathCode">$l_2$pan>pan>pan>) distance to learn the dictionary and encode the histograms. However, it has been shown that Histogram Intersection Kernel (HIK) is more effective to compare histograms. In this paper, we combine Histogram Intersection Kernel with local sparse coding. We implement dictionary learning and feature encoding on the mapping space that corresponds to the kernel. To encode the features, we propose two methods: one accurate method generating codes consisting of positive and negative values and one approximate method generating only non-negative values. Both of the two encoding methods run very fast. To verify our method, we conduct some experiments on two popular datasets: Caltech-101 and Caltech-256. The results show that the features extracted by our method are more discriminative than other methods and it reaches state-of-the-art result on Caltech-101 when taking single descriptor HOG as input. In addition, it shows that the codes with non-negative constraint are more effective than that without the constraint.