International Journal of Science, Engineering and TechnologyAuthors: C.Sai Kalyani Deepthi, G.Dhana Lakshmi, SK. Najmabi, B.Lakshmi Mounika, K. Srividhya
Abstract: Quickly growing the microplastic contamination of the water body has emerged as a great ecological menace, compromising the marine biodiversity and water quality. Obvious limitations of traditionally used methods in detection, such as CNN-based models and other methods based on machine learning, include high computational cost, suboptimal accuracy in complex visual scenarios, sensitivity to the environment, and restrictions on the use in practice in real time. In order to overcome these drawbacks, this study suggests a Vision Transformer (ViT)-based architecture to effectively and precisely detect and classify microplastics on water images. The procedure starts with the further image enhancement based on Contrast-Limited Adaptive Histogram Equalization (CLAHE) that enhances the visibility of microplastics and reduces the noise. Images are split into patches and treated with transformer encoder layers with the help of multi-head self-attention to extract the global contextual information efficiently. A ViT-based decoder allows accurate classifying and segmenting of microplastics by type and size, which is trained on a hybrid loss based on cross-entropy and Dice losses to maximize pixel-wise accuracy. Experimental outcomes show that the proposed ViT model is better than traditional TinyML and CNN-based models, its detection accuracy is over 97, feature extraction accuracy is over 95, and its precision is 96.5, and the encoding time is cut by about 30. The model has strong generalization capabilities on a wide range of aquatic data with uniform training-validation results, which make it applicable to environmental monitoring. Overall, this ViT-based solution is scalable, computationally efficient, and highly accurate when it comes to assessing microplastic pollution, which enhances the conservation of the ecological environment, and offers a solution to real-time monitoring of the aquatic environment.
DOI: http://doi.org/10.5281/zenodo.20390106
