Contextual understanding on natural scene images for improved annotation using heuristics and methods

Abstract

Advancements in object or image annotation techniques have revolutionized computer vision and image analysis applications. By accurately labeling and marking objects or regions of interest within images, these advancements enable more efficient and accurate understanding of visual data in Natural Scene Images (NSI). Object detection and recognition have greatly benefited from improved annotation techniques, resulting in enhanced algorithms for identifying and classifying objects. Image annotation has played a vital role in training models for image captioning and understanding, as well as in augmented reality and virtual reality applications where virtual content needs to align seamlessly with the real world. In general, advancements in object or image annotation have greatly enhanced the accuracy and efficiency of computer vision algorithms, impacting diverse fields such as healthcare, transportation, assisting visually impaired, entertainment, and beyond. newlineFirst, this work presents an object detection scheme that utilizes the AlexNet deep learning model as its base. Different optimizers including SGDM, RMSProp, and Adam are employed during the execution of the AlexNet model, with performance evaluation conducted on the Flicker dataset. The results demonstrate that the Adam optimizer outperforms the others. In addition to object detection, this work introduces a context-based Hidden Markov Model (HIV IM) for improving the Image annotation based on heuristic attributes of objects and their inter-relationships. The HMM model enhances the understanding of the objects by generating annotation for the image that optimally matches the synonymous captions present in the dataset. newline