A hybrid attention-based deep learning framework for high-accuracy microplastic detection and monitoring in water environments

Microplastic pollution has emerged as a major environmental concern, necessitating accurate and efficient monitoring technologies for aquatic ecosystems. While deep learning-based image analysis has shown promise for automated microplastic detection, existing methods often suffer from limited robustness, insufficient validation under realistic environmental conditions, and reduced performance in complex backgrounds. To address these challenges, this study proposes Yolov7CS, a hybrid attention-enhanced object detection framework for microplastic identification and monitoring. The proposed model integrates a Convolutional Block Attention Module (CBAM), combining channel and spatial attention mechanisms, into the Yolov7 architecture to improve feature extraction and discrimination of diverse microplastic morphologies. A high-resolution image dataset containing 4260 images across seven particle categories was established under multiple background conditions. The dataset comprised six microplastic categories-Film, Fiber, Foam, Fragment, Pellet, and Tire-and one non-microplastic interference category, Road Salt. In addition, a real-world water circulation platform was developed to evaluate model performance under practical monitoring scenarios involving water interference, illumination variation, and background complexity. Among the evaluated models, Yolov7CS achieved the highest overall Precision, mAP@50-95, and F1-score, while tying for the highest Recall and mAP@50. These results indicate competitive and relatively balanced performance under the evaluated conditions.