A lightweight convolutional neural network for large-scale Chinese image caption

Abstract

Image caption is a high-level task in the area of image understanding, in which most of the models adopt a convolutional neural network (CNN) to extract image features assigning a recurrent neural network (RNN) to generate sentences. Researchers tend to design complex networks with deeper layers to improve the performance of feature extraction in recent years. Increasing the size of the network could obtain features of high quality, but it is not an efficient way in terms of computational cost. A large number of parameters brought by CNN makes the research difficult to apply in human daily life. In order to reduce the information loss of the convolutional process with less cost, we propose a lightweight convolutional neural network, named as Bifurcate-CNN (B-CNN). Furthermore, recent works are devoted to generating captions in English, in this paper, we develop an image caption model that generates descriptions in Chinese. Compared with Inception-v3, the depth of our model is shallower with fewer parameters, and the computational cost is lower. Evaluated on the AI CHALLENGER dataset, we prove that our model can enhance the performance, improving BLEU-4 from 46.1 to 49.9 and CIDEr from 142.5 to 156.6 respectively.