【摘要】 研究了基于层次形状特征提取的图像分类。针对从初级视觉皮层(V1)提取的条形特征对目标形状的描述不充分,提出了一种层次模型(V1-V2-V4),以进一步提取角形、曲率特征。模型中V1层的条形特征提取采用Gabor模拟;V2层结合了抑制噪声的3D高斯差分(DOG),并使用滤波方向相差90度的Gabor滤波器组提取多尺度角形特征;V4层通过曲率域计算来描述目标轮廓的形变程度,并最终提取融合曲率与梯度方向的直方图特征。该模型的优势在于,通过角形以及曲率计算的层次表达,可有效增强目标形状的关键特征点(如角点位置)的提取,并且结合曲率与梯度的直方图描述,也可有效弥补单一曲率或梯度特征局部描述不足的问题。在MNIST手写数字与21类遥感影像上的实验表明,曲率与梯度的融合特征的运用可获得98.94%的数字识别精度,同时在遥感影像分类中也可获得较好的分类效果。更多还原

【Abstract】 The image classification based on hierarchical shape feature extraction was studied.Considering that the bar features extracted from the primary visual cortex of V1 can not sufficiently represent the shape of an object,a new hierarchical model of V1-V2-V4 was presented to further extract the features of angle and curvature.Under the model,the Gabor function is adopted to simulate V1 to extract the bar features,the noisy inhibition using 3D-D0G(difference of Gaussian) combined with a group of Gabor filters with the filtering direction difference of 90°,is introduced to extract the multi-scale angle features in V2,and in V4,the deformation of the shape of an object is described through the computation of the curvature field,to finally extract the histogram features fusing the curvature and gradient orientation.The advantage of the proposed model is that the hierarchical representation based on the computation of angle and curvature helps to extract the key points of object shape such as the corners.Moreover,the use of fusion features also remedies the inadequacy of the object description only using curvature or gradient feature..The experiments on MNIST handwritten digits and 21 remote sensing images demonstrated that the use of the features fusing curvature and gradient orientation achieved the digit recognition accuracy of 98.94%,and obtained the better result in classification of remote sensing images.更多还原