Dominant orientation patch matching for HMAX

详细信息    查看全文

• 作者：Yan-Feng Lu^a ; ^{yanfeng.lv@ia.ac.cn" class="auth_mail" title="E-mail the corresponding author}Author Vitae ; Hua-Zhen Zhang^b ; ^{zhanghz@korea.ac.kr" class="auth_mail" title="E-mail the corresponding author}Author Vitae ; Tae-Koo Kang^c ; ^{tkkang@smu.ac.kr" class="auth_mail" title="E-mail the corresponding author}Author Vitae ; Myo-Taeg Lim^b ; ^{mlim@korea.ac.kr" class="auth_mail" title="E-mail the corresponding author}Author Vitae
• 关键词：Object recognition ; Classification ; HMAX ; Dominant orientation ; Patch ; Matching
• 刊名：Neurocomputing
• 出版年：2016
• 出版时间：12 June 2016
• 年：2016
• 卷：193
• 期：Complete
• 页码：155-166
• 全文大小：9039 K

文摘

The biologically inspired model for object recognition, Hierarchical Model and X (HMAX), has attracted considerable attention in recent years. HMAX is robust (i.e., shift- and scale-invariant), but it is sensitive to rotational deformation, which greatly limits its performance in object recognition. The main reason for this is that HMAX lacks an appropriate directional module against rotational deformation, thereby often leading to mismatch. To address this issue, we propose a novel patch-matching method for HMAX called Dominant Orientation Patch Matching (DOPM), which calculates the dominant orientation of the selected patches and implements patch-to-patch matching. In contrast to patch matching with the whole target image (second layer C1) in the conventional HMAX model, which involves huge amounts of redundant information in the feature representation, the DOPM-based HMAX model (D-HMAX) quantizes the C1 layer to patch sets with better distinctiveness, then realizes patch-to-patch matching based on the dominant orientation. To show the effectiveness of D-HMAX, we apply it to object categorization and conduct experiments on the CalTech101, CalTech05, GRAZ01, and GRAZ02 databases. Our experimental results demonstrate that D-HMAX outperforms conventional HMAX and is comparable to existing architectures that have a similar framework.