Abstract: Because of its high accuracy in image recognition, convolutional neural network is very popular in the field of image retrieval. However, in the face of large datasets, the high deep feature dimension extracted by convolutional neural network is likely to cause “the curse of dimensionality”. Aiming at the problem of the high dimension of deep features in image retrieval, a novel image retrieval algorithm based on adaptive fusion network feature extraction with hash feature reduction is proposed. Due to the high complexity of high-dimensional features in traditional hash processing, an adaptive fusion module is added to the convolutional neural network to re-integrate the features, to enhance the capability of feature representation and reduce the feature dimension. Then, the deep feature dimension is reduced by feature sparsity optimization algorithm for the second time, and the reduced hashing code is obtained by mapping. Finally, using Inception network as the basic model, a number of experiments were conducted on CIFAR-10 and ImageNet datasets. Experimental results show that this algorithm can effectively improve the efficiency of image retrieval.

Key words: convolutional neural network, deep feature, feature dimensionality reduction