Image Caption Generation using Deep Learning Approaches

Image caption generation is an endeavor to provide a meaningful textual description that involves an image. Extracted information is relevant to the activities present in the images. ResNet (Residual Network) is well known for its ability to classify images, having developed deep hierarchical representations. The intention of this paper is to use ResNet with various smart filters to classify images more deeply, enabling the generation of genuine and meaningful descriptions that are highly precise with respect to the reference captions. Here, the work uses a smart filtering technique to enhance images, a CNN to encode features, model training, and thereafter an RNN (Recurrent Neural Network) to decode the features. ResNet is a very effective model for computer vision tasks, especially object classification and semantic analysis. ResNet is well known for residual connections, which are also known as skipping connections that solve the vanishing gradient problem, which is a crucial problem in deep learning. Here, the MSCOCO (Microsoft Common Object in Context) benchmark is used to train the model, which is a large dataset with reference annotations useful for various computer vision tasks. ResNet helps enhance generalization capability, which is particularly useful for diverse images. As per the results obtained, BLUE scores are B1: 0.579, B2: 0.404, B3: 0.279, B4: 0.191; METEOR: 0.195; ROUGE: 0.396; and CIDEr: 0.6.