Precise and Faster Image Description Generation with Limited Resources Using an Improved Hybrid Deep Model

We propose a model that performs image captioning efficiently based on entity relations, followed by a deep learning-based encoder and decoder model. In order to make image captioning more precise, the proposed model uses Inception-Resnet(version-2) as an encoder and GRU as a decoder. To develop a less expensive and effective image captioning model in accordance with accelerating the training process by reducing the effect of vanishing gradient issues, residual connections are introduced in Inception architecture. Furthermore, the effectiveness of the proposed model has been significantly enhanced by associating the Bahadanu Attention model with GRU. To cut down the computation time and make it a less resource-consuming captioning model, a compact form of the vocabulary of informative words is taken into consideration. The proposed work makes use of the convolution base of the hybrid model to start learning alignment from scratch and learn the correlation among different images and descriptions. The proposed image text generation model is evaluated on Flickr 8k, Flickr 30k, and MSCOCO datasets, and thereby, it produces convincing results on assessments.