胶囊网络下基于三元损失的图像相似性学习

Image Similarity Learning Based on Triple Loss under Capsule Network

学习细粒度图像相似性是机器学习中一项富有挑战性的任务,尤其是相似性很高的同类图像.本文提出了基于三元损失函数和胶囊网络的图像相似性学习方法,三元损失能够反映图像之间的"相对"相似性,而胶囊网络比普通深度学习网络具有更优的图像特征提取能力,将两者有机结合在一起形成的新型网络能够很好地完成图像相似度的学习任务.实验结果表明,本文提出的新方法相比于一般的深度学习方法,准确率有很大地提高.

Learning fine-grained image similarity is a challenging task in machine learning, especially for highly similar images. We establish an image similarity learning method based on triplet loss function and capsule network. The triplet loss can reflect the "relative" similarity between images, while the capsule network has better image feature extraction ability than the ordinary deep learning network. The new network combining them together can well accomplish the task of learning image similarity. The experimental results show that the proposed method has a higher accuracy than ordinary deep learning method.