In recent years, deep learning has become a key technology for image classification and feature extraction, offering new possibilities for analyzing complex visual data. In this paper, we design a hybrid framework that joins the strengths of the Swin Transformer and EfficientNet models to improve the detection and classification of marine plastic waste. The combined model takes advantage of the Swin Transformer’s ability to capture multi-scale contextual details and EfficientNet’s lightweight design for faster computation. The training was performed using the PyTorch framework with mixed-precision optimization, allowing efficient use of GPU memory. The dataset was preprocessed using normalization and basic augmentations such as flips and rotations to make the model more robust. During the training stage, we applied gradient scaling and automated mixed-precision (AMP) to stabilize the process and accelerate convergence. The final trained model was exported in a portable format for easy deployment. The results indicate that the proposed hybrid setup achieves better accuracy and computational efficiency than standalone models, showing potential for use in large-scale environmental monitoring tasks.