Pre-trained CNNs as Feature-Extraction Modules for Image Captioning

In this work, we present a thorough experimental study about feature extraction using Convolutional NeuralNetworks (CNNs) for the task of image captioning in context deep learning. We perform set 72experiments on 12 classification CNNs pre-trained ImageNet [29] dataset. The features areextracted from last layer after removing fully connected and fed into model. usea unified model with fixed vocabulary size across all experiments to effect changingthe CNN extractor quality. scores are calculated standard metrics inimage captioning. find strong relationship between structure datasetand prove that VGG models give least quality among testedCNNs. end, recommend each evaluation metricswe want optimise, show connection our results previous works. To knowledge, thiswork is most comprehensive comparison extractors