Holistic Scene Recognition through U-Net Semantic Segmentation and CNN

Abstract

The rapid pace at which technology is developing has resulted in a significant rise in the intelligence of machines. These days, researchers are dedicated on giving machines human-like thought processes and reasoning abilities. There is still an immense gap in current machines' comprehension of real-world scenes, regardless the capacity to gather and evaluate input via sensors. Understanding scenes has been an essential field of research. In order to cope with this issue, we have proposed a model that leverages depth information to enable real-time scene perception and comprehension for robots, comparable to that of humans. By giving machines human-like scenario interpretation abilities, this creative approach aspires to ease the way for ground-breaking advancements in a wide range of applications. The suggested recognition technique provides a novel segmentation strategy that leverages deep learning to overcome the challenge of scene recognition. By utilizing a U-Net deep learning model, the system is able to perform robust multi-object segmentation after learning an entire scene model. In order to recognize scenes, unique features based on entropy are gathered from the divided objects and input into a CNN classifier. This integrated strategy, which integrates CNN-based classification and deep learning segmentation with unique feature engineering, has demonstrated significant improvements over the state-of-the-art systems currently in use.