Abstract: To solve the problem that low-level features extracted by unsupervised learning methods are easily disturbed by image’s rotation and scaling as well as difficult to distinguish when used in feature description, a local feature description algorithm is proposed based on improved Convolutional Auto-Encode（CAE-D）. Evaluating the convolution kernel’s performance by information entropy, a convolution kernel’s entropy constraint rule is proposed to improve the distinguish ability of convolution feature description through convolution kernels carrying local information. Traditional SIFT’s orientation assignment algorithm is used to assign the main direction of local image before feature description, and the feature-map is down-sampled to enhance rotation-invariance and robustness of the feature description. The results of image matching show that CAE-D is competitive with the performance of KAZE and SIFT descriptor in geometric and photometric deformations and takes 47. 14% less time than SIFT.

Key words: unsupervised learning, feature description, convolutional auto-encoder, information entropy

摘要： 针对非监督学习方法提取的底层特征用于特征描述时可区分性不强，对图像旋转、尺度等变换敏感的问题，提出了一种改进卷积自编码器的局部特征描述算法（Convolutional Auto-Encoder Descriptor，CAE-D）。CAE-D算法利用信息熵评价卷积核性能，提出在CAE中添加卷积核信息熵约束规则，通过均值化卷积核携带的局部特征信息，提升卷积特征描述的可区分性；在特征描述前使用传统SIFT中主方向分配算法确定局部图像的主方向，并引入降采样操作，进一步提升特征描述的旋转不变性及鲁棒性。图像匹配实验结果验证了改进策略的有效性，CAE-D算法优于当前先进的KAZE、SIFT，而运行时间相比SIFT缩短了47.14%。

关键词: 非监督学习, 特征描述, 卷积自编码器, 信息熵