摘要
图像中物体的轮廓是稳定的,是旋转、平移和尺度等不变性识别的基础.本文对初级视皮层的神经元细胞进行建模,借鉴神经元方位选择性提取图片朝向信息,实现更接近大脑感知的形状特征表示策略.首先,以线段为基元表征物体的几何特征,引入朝向强度与尺度空间来获取更丰富的轮廓信息;其次,利用神经编码稀疏性指标,改进朝向关系匹配方法和尺度融合策略,获得了表示能力更强的图像形状特征.实验表明,基于该特征表示策略的形状检索算法实现了更高的检索正确率,在存在遮挡和噪声干扰情况下,本文方法表现出更好的鲁棒性.
As a stable feature of the objects in images,the contour is one of the bases of invariant recognition,such as rotation,translation,and scale.Inspired by the orientation selectivity in the primary visual cortex,this paper extracts the orientation intensities in images by modeling the receptive field of visual neurons and proposes a shape feature representation strategy similar to brain perception.First,to obtain abundant contour information,the primitive geometric features of the object are represented by the orientation intensity of line segments at multiple scales.Second,we improve the orientation matching and multiscale fusion strategy by using the sparsity index.We also obtain more representative shape features.The experimental results suggest that the shape retrieval algorithm based on our feature representation strategy achieves a higher retrieval accuracy and stronger robustness in the presence of occlusion and noise.
作者
段婷
李明
沈辉
胡德文
DUAN Ting;LI Ming;SHEN Hui;HU DeWen(College of Intelligence Science and Technology,National University of Defense Technology,Changsha 410073,China)
出处
《中国科学：技术科学》
EI
CSCD
北大核心
2022年第5期682-688,共7页
Scientia Sinica(Technologica)
基金
国家重点研发计划(编号:2018YFB1305101)
国家自然科学基金(批准号:62076248,62036013)资助项目。
关键词
形状特征表示
多尺度融合
方位选择性
几何特征提取
图像表征
shape feature representation
multiscale fusion
orientation selectivity
geometric feature extraction
image representation
