基于点击数据的图像识别

Image recognition based on click data

随着细粒度图像分类研究的不断深入,用户点击数据逐渐被人们当成可靠的语义特征。由于用户点击数据集规模巨大且存在大量冗余,直接使用点击特征进行识别也存在诸多挑战。该文提出利用文本聚类降低文本空间并优化原始点击特征,从而建立精简的文本空间来表征图像,该方法能更好地合并语义相近的文本。在微软发布的Clickture-Dog大数据集上进行的大量实验表明,点击向量特征优于传统图像的视觉特征,图像识别任务中的准确率也更高;基于视觉相似度的传播算法能帮助提高点击特征的表征能力;在大规模文本聚类中,基于稀疏编码的聚类方式识别率达到了58.24%。

For fine⁃grained image classification and recognition,users’click information are proved to be useful for construct image semantic features.With user⁃click data each image is represented as query⁃click⁃frequency vector.Compared with traditional visual features.However,due to the redundancy and huge size of the text set,there are also many challenges in using the click feature directly for classification and recognition.This paper proposes to use text clustering to reduce the text space and optimize the original click feature,so as to establish compact and effective text space to represent the image.This method can better merge semantically similar text.Extensive experiments have been carried out on Clickture⁃Dog dataset.Experimental results show that in image representation,click vector feature is superior to traditional image visual feature,and the accuracy of image recognition task is higher;The propagation algorithm based on visual similarity can help improve the representations of click features;In the large text clustering,the accuracy of clustering method based on sparse coding has reached 58.24%.