Image captioning models generate textual descriptions for given images, by combining two domains - computer vision and natural language processing. Topic-based image captioning models utilize topics as auxiliary guiding information to generate semantically rich captions. The state-of-the-art techniques apply attention mechanisms on top of these topic-based models; thus any improvement in the underlying models will enhance the overall performance of the captioning process. The current research work proposes an efficient image captioning model with a reduced number of parameters compared to its base variant algorithm. The refinements are applied in two contexts. The feature extraction phase is implemented using EfficientNet, which produces a feature vector of size 1280. This improvement has significantly reduced the feature vector of size 2048 when used with an Inception-V3 model. Also, lower dimensional feature embedding using GloVe-100 scheme is incorporated. The proposed model is trained using two datasets Flickr8k of 8091 images and Flickr30K with 31800 images. The experimental results show that the number of model parameters has been educed by 14.146%. Consequently, this model with the proposed replacements is more preferable over the base model. Additionally, the proposed model outperformed topic-specific caption generation in terms of both BLEU and METEOR scores.