Abstract: To address the issues of low efficiency caused by the long imaging time of scanning electron microscopes in deep learning-based detection methods, the difficulty in balancing real-time performance, detection efficiency, and accuracy in detection algorithms, and the lack of a scientific deep learning dataset suitable for optical microscope images in microplastic detection, the YOLOv8-REM algorithm is proposed, which is an improvement based on the YOLOv8s framework, and constructs a scientific optical microscope image dataset for microplastics. In the algorithm improvement, firstly, by reducing the number of layers in the YOLOv8s backbone network, the loss of small-object information due to deep convolutions is minimized, preventing small-object features from being overshadowed by deep-layer features. This allows the network to focus more on shallow features, thus enhancing both small-object detection capability and processing speed. Additionally, an efficient multi-scale attention module is introduced, which uses exponential weighting to perform weighted averaging of feature maps, emphasizing important features and suppressing irrelevant ones. This gradually updates the representation of the feature map, enabling the model to better capture long-term dependencies and dynamic changes, increasing its focus on key information and enhancing its ability to detect small-object features. Finally, a multi-scale feature fusion module is designed to supplement small-object information by integrating features from multiple scales, further optimizing detection performance. Experimental results show that the YOLOv8-REM algorithm achieves an average detection precision (mAP@0.5:0.95) of 78.4%, a 5.2 percentage point improvement over the original YOLOv8s, and a 27.03% reduction in parameter count. While ensuring real-time performance, this method achieves efficient and accurate detection of microplastics in water based on optical microscope images.