The ubiquitous presence of microplastics across environmental compartments presents a formidable ecotoxicological and risk assessment challenge, fundamentally complicated by the link between microplastic morphology and differential toxicological outcomes. Current analytical methods face a significant measurement bottleneck, hindering the precise, high-throughput characterization needed for robust mechanistic and exposure studies. To address this, we introduce MNv4-Conv-M-fpn, a novel deep learning model specifically engineered for multi-class microplastic segmentation and morphological characterization from microscopic images. This model is designed to provide the toxicologically-relevant granularity required for rigorous risk assessment, segmenting images into six classes: five distinct microplastic categories (fiber, fragment, sphere, foam, and film) and the background. By incorporating advanced architectural features—including transfer learning, a Feature Pyramid Network, and a Feature Fusion Module—our approach achieves high accuracy, computational efficiency, and near real-time inference speed. Comprehensive validation using a diverse dataset demonstrates that MNv4-Conv-M-fpn outperforms existing segmentation methods while maintaining low computational load. This makes the model well-suited for high-throughput deployment in environmental laboratories and resource-constrained monitoring efforts. This approach offers a valuable tool for environmental monitoring, enabling more precise and scalable analysis of microplastic pollution in various ecosystems.
The lower quality of frames in satellite videos compared to natural videos poses significant challenges in capturing detailed information for alignment and fusion in the image space. In this paper, we introduce a novel frequency-aware enhancement network (FAENet) for satellite video super-resolution (SVSR), which tackles these challenges from a frequency-domain perspective. By leveraging frequency components, FAENet amplifies the distinctions between frames and between objects, thereby improving alignment and reconstruction accuracy. Firstly, the proposed Frequency Alignment Compensation Mechanism (FACM) incorporates a frequency-domain distribution alignment function to enable effective alignment compensation. This mechanism can be seamlessly integrated into existing alignment methods designed for natural video, thereby enhancing their applicability to SVSR tasks. Secondly, we introduce the Frequency Prompt Enhancement Block (FPEB), which facilitates edge reconstruction by leveraging frequency-domain prompts to distinguish objects from artifacts, thereby improving the clarity and accuracy of reconstructed edges. The proposed FAENet achieves 35.33 dB PSNR on the Jilin-189 dataset and 40.57 dB on the SAT-MTB-VSR dataset, outperforming other state-of-the-art compared methods and demonstrating its effectiveness and robustness in addressing the unique challenges of SVSR.