Scene understanding refers to the ability of a computer system to analyze and interpret a given scene, by extracting meaningful information from visual elements, such as images or videos. The objective is to achieve a higher level of comprehension and enumeration of the various objects present in the scene, their attributes, the visual relationships between them, and then expressing such understanding by providing an accurate and complete description of the given scene in natural language. Scene understanding plays a crucial role in enabling machines to perceive, analyze, and interpret their environments, empowering them to make intelligent decisions and facilitate natural and intuitive communication with humans. This has significant implications for the advancement of scientific research that relies heavily on visual data, particularly in medicine, industry, and astronomy. Several techniques and methodologies have been developed to achieve a deeper understanding of the environment, the best known and commonly used are Deep learning models, particularly CNNs, GNNs, and neural language generation models, such as (RNNs) or transformer-based models like GPT.As the volume of visual data continues to grow, the number and diversity of visual relationships between objects also increase, which presents a significant challenge in discovering and quantifying these relationships, taking into consideration the complexity of the used techniques and the available material resources. With the aim of addressing this issue, our research focuses on exploring alternative methods that can enhance the visual relationship detection accuracy, while reducing the computational time using available material rather than acquiring more efficient ones. Our approach involves comparing two CNN models, namely resnet101 and resnet152, to determine their effectiveness in enhancing visual relationship detection.