基于半监督深度网络学习的细粒度图像检索

Fine-graining image retrieval based on semi-supervised training method for deep network

针对传统图像特征在细粒度图像检索上的不足,提出一种基于稀疏编码的半监督深度学习算法,来提高细粒度图像的检索性能。该算法首先通过少量带标签的训练样本来微调深度神经网络模型,然后通过该模型生成其对应的稀疏编码,最后以重构的稀疏编码作为网络训练的监督信息,通过最小化交叉熵损失函数来优化深度神经网络模型。在公开的细粒度图像数据集上进行了对比实验,结果表明,所提出的算法在少量标记样本的情况下具有较好的检索性能。

In view of the shortcomings of traditional image features in fine-grained image retrieval, a semi-supervised deep learning algorithm based on sparse coding is proposed to improve the retrieval performance for fine-grained images. The algorithm first fine-tunes the deep network model by a small number of trained samples with labels. Then, the corresponding sparse codes are generated through the model. Finally, the reconstructed sparse codes are served as the supervised information of network training to optimize the deep neural network model by minimizing the cross entropy loss function. A comparative experiment is performed on the public fine-grained image dataset, showing that the algorithm achieves better retrieval performance in the case of a few labeled samples.