YOLO-CSM: an enhanced YOLOv11n-based framework for efficient microplastic detection in environmental monitoring

As an emerging environmental pollutant, the potential toxic effects of microplastics have become a topic of global concern. To address the issues with traditional detection methods- such as cumbersome operations, time-consuming processes, and reliance on human expertise-this study proposes an improved model based on YOLOv11n, named YOLO-CSM, for the efficient detection and classification of microplastics. First, partial multiscale convolutional modules (CSP-P) are introduced into the backbone and neck networks. By utilizing multiscale convolutional kernels and a cross-stage feature fusion mechanism, the model's ability to extract fine-grained features and detect minute objects is enhanced. Second, an SKFP module based on shared and dilated convolutions is designed to expand the receptive field and improve multiscale feature representation capabilities. Finally, the MSR module in the detection head dynamically adjusts feature weights via an adaptive average pooling layer and combines this with fully connected layers to strengthen cross-scale feature fusion, thereby improving the model's ability to recognize targets in complex backgrounds. Experimental results show that the improved model achieves mAP50 and mAP50-95 values of 92.06% and 44.08%, respectively. The complete model required 8.7 GFLOPs and achieved an inference speed of 143.85 FPS, indicating that it retained real-time detection capability despite the increase in computational complexity. Its overall performance surpasses that of the baseline YOLOv11n and other mainstream models, providing an efficient and reliable visual method for microplastic detection.