In today's rapidly developing digital technologies, image processing occupies a key place in the fields of computer vision and graphics. One of the most popular tasks is the automatic separation of foreground objects from the background, which is of critical importance for web design, e-commerce, the advertising industry, and augmented reality systems. Traditional manual editing methods are laborious and dependent on user experience, especially when working with complex textures or fuzzy object boundaries. The development of deep learning methods make it possible to automate these processes, providing high accuracy and speed of processing in real time. The aim of the work is to analyze existing algorithms for automatic background removal and develop an effective methodology for object segmentation using modern computer vision models to improve the quality of multimedia content. The work uses a comprehensive approach based on the use of the hybrid neural network architecture BEN2. The key feature of the method is the implementation of an innovative Confidence Guided Matting pipeline, which implements a two-stage Bayesian approach: first, coarse segmentation (BEN Base) is performed to create a "draft" mask, and then targeted refinement of boundaries in areas of low confidence of the model (BEN Refiner). As part of the research, a software application based on the Streamlit framework was developed, which provides automated inference of the BEN2 model. The system supports loading images in popular graphic formats, batch data processing, and visualization of results using an interactive slider. The high efficiency of the algorithm was experimentally confirmed: for complex objects, the architecture provides accurate matting of hair and small details, minimizing artifacts such as "ragged edges". The processing workflow was optimized by using local caching of models and supporting acceleration on GPU/CPU. The proposed approach based on the BEN2 architecture and boundary refinement mechanisms demonstrated higher accuracy compared to classical one-stage segmentation methods. The developed system is scalable and suitable for integration into real multimedia processing information systems, which allows significantly reducing the time for preparing graphic content and increasing its professional level.