基于轻量级深度哈希网络的细粒度图像检索

Fine-grained Image Retrieval Based on Lightweight Deep Hash Network

深度哈希利用端到端的框架同时进行特征学习及哈希编码,两者相互促进,提取到合适的特征以及生成优质的哈希码。然而,深度哈希方法在图像检索研究中仍面临一些挑战:(1)大多数现有的深度哈希方法使用复杂的神经网络作为基础网络,网络参数增多,模型越来越大,在一些移动端和嵌入式设备上很难写入几十甚至上百MB的模型。(2)目前,大多数深度哈希方法使用具有高时间复杂度的损失函数来保留原始数据空间和哈希编码之间的相似性,无法在时间和准确性上实现双赢。针对上述问题,文中利用轻量级网络作为主干网络,并针对细粒度图像类内距离大、类间距离小的特点,提出跨层的多尺度Non-Local模块进行特征融合。其次,在分类层之前加入哈希编码层,同时利用简单且有效的交叉熵损失代替复杂的成对相似性保留损失。实验结果证明,在三个公开的细粒度图像数据集上,与其他先进的图像检索算法相比,文中提出的方法在检索性能上具有明显的优势,其top1的检索精度均可达80%以上,且超出第二名10%以上。

Deep hash uses an end-to-end framework to perform feature learning and hash coding at the same time,and the two promote each other to extract appropriate features and generate high-quality hash codes.However,deep hashing methods still face some challenges in image retrieval research.(1)Most of the existing deep hashing methods use complex neural networks as the basic network,with more network parameters and larger models,which is difficult to write tens or even hundreds of MB of models on some mobile terminals and embedded devices.(2)At present,most deep hashing methods use loss functions with high time complexity to preserve the similarity between the original data space and the hash coding space,and cannot achieve a win-win situation in terms of time and accuracy.In response to the above problems,we use a lightweight network as the backbone network and propose a cross-layer multi-scale feature fusion for the characteristics of fine-grained images with large intra-class distances and small inter-class distances.Secondly,a hash coding layer is added before the classification layer,and a simple and effective cross-entropy loss is used to replace the complex pairwise similarity preservation loss.The experiment proves that compared with other advanced image retrieval algorithms,the proposed method has obvious advantages in retrieval performance on three publicly available fine-grained image datasets,and its top1 retrieval accuracy can reach more than 80%,and exceed the second place by more than 10%.