Logo image
Advancing image captioning using custom cnn-rnn architectures with attention mechanisms, commonsense reasoning and advance data visualization
Thesis   Open access

Advancing image captioning using custom cnn-rnn architectures with attention mechanisms, commonsense reasoning and advance data visualization

Nikhil Teja Putta
California State University, Sacramento
Master of Science (MS), California State University, Sacramento
06/09/2026
Handle:
https://hdl.handle.net/20.500.12741/rep:14146

Abstract

Image captioning systems have become a significant part of computer vision applications and also game changing part of retail-analytics, automobiles, defense, agriculture and healthcare applications. The project emphasizes image captioning systems that leverage advanced technologies, redefining how the images are understood, interpreted and captioned. Early image captioning systems were mostly fixed-length vector-based models using convolutional neural networks (CNNs). This mechanism led to loss of fine -grained spatial details resulting in generic, less accurate captions, as these models did not improve the insights grounded in the real-world contexts. The model utilizes Flickr data, processed and cleaned for efficient usage. The dataset consists of 30k images and 150k captions further simplified into 10k images and 30k captions. By employing a hybrid CNN-RNN model with enhanced attention mechanisms, the system can precisely focus model has a clear idea where to focus precisely girds and pixels improving the cohesion and contextual precision of captions. The integration of logical reasoning, common sense knowledge ensures the captions are semantically rich and follow real world scenarios. Overall, incorporating of VIS modules such as commonsense reasoning score and heatmaps not only boosts performance but also clarifies the system decision making. The entire performance is calculated by the commonsense reasoning score and the concentration of the model on the image, as highlighted through heatmap. With these enhancements, the model is well enhanced now it is ready to pave way for more interactive and interpretable image captioning applications.
pdf
NikhilTejaPutta_Spring2051.88 MBDownloadView
Text Project Open Access

Metrics

1 Record Views

Details

Logo image