We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Microplastic Image Segmentation for Edge Devices using Lightweight U-Net
Summary
Researchers developed a lightweight U-Net model for microplastic image segmentation deployable on portable edge devices, testing depthwise separable convolution layers to maintain detection accuracy while reducing computational load. Using depthwise separable convolutions achieved similar mean intersection-over-union performance to the standard U-Net while reducing operations by 82.6%.
해양 생태계에 부적절한 영향을 끼치는 미세플라스틱을 검출하여 배출되는 양을 줄이는 작업이 필요하다. 미세플라스틱은 현미경 이미지 분할 딥러닝 모델을 적용해 검출할 수 있는데, 정확도를 높이기 위해서는 데이터의 다운샘플링을 지양하는 것이 중요하다. 하지만, 이를 위해서는 일반적으로 모델의 파라미터 및 연산량이 증가하는 문제가 있어 휴대형 장치에 실시간으로 동작하기 어렵다. 따라서 딥러닝 모델을 휴대형 장치를 사용해 현장에서 채취한 시료로부터 즉각 미세플라스틱을 검출하는 과정에는 적용하기 위해서는 데이터의 해상도를 유지하되 경량화 된 모델이 필요하다. 모델 경량화 실험을 위해서 이미지 분할 분야의 대표적인 모델인 U-Net의 합성곱 계층에 기존의 다양한 경량화 방법을 적용하여 U-Net과의 성능을 비교하였다. Depthwise Separable 합성곱 계층을 사용한 경우 U-Net에 비하여 유사한 mIoU 성능을 보이며, 82.6% 감소된 연산량을 보였다. 또 행렬 분해 및 Group 합성곱을 적용한 경우는 U-Net에 비해 7.41%의 mIoU 성능 하락이 있었지만, 99.55% 감소된 연산량을 보인다. 이는 현장에서 즉각적으로 데이터를 취득해야 하는 휴대형 장치에 적용해 이점을 얻을 수 있을 것으로 생각된다.
Sign in to start a discussion.
More Papers Like This
Detection of Microplastics in Coastal Environments Based on Semantic Segmentation
Researchers developed a deep learning semantic segmentation approach for detecting microplastics on sandy beaches at the pixel level, evaluating 12 models including U-Net variants and transformer architectures under real-world conditions.
Slim Deep Learning Approach for Microplastics Image Classification in the Marine Environment
Researchers developed a lightweight convolutional neural network called the Slim-DL-Model for classifying microplastics in marine environment images, designed to overcome the computational demands of existing architectures like VGG16 and ResNet for real-time field applications. The model achieves competitive classification accuracy while significantly reducing computational requirements, enabling deployable microplastic monitoring systems.
Deep learning approach for automatic microplastics counting and classification
Researchers developed a deep learning architecture combining U-Net segmentation and VGG16 classification to automatically count and categorise microplastic particles of 1-5 mm into fragments, pellets, and lines from digital camera images. The system reduces the cost and time of traditional microplastic quantification methods while enabling high-throughput monitoring.
MP-Net: Deep learning-based segmentation for fluorescence microscopy images of microplastics isolated from clams
Researchers developed MP-Net, a deep learning model based on U-Net architecture, that accurately segments and quantifies fluorescent microplastics in microscopy images of clams, achieving over 90% accuracy and enabling faster, more reliable environmental monitoring.
A Deep Learning Model for Automatic Plastic Mapping Using Unmanned Aerial Vehicle (UAV) Data
Researchers applied a deep learning semantic segmentation model (ResUNet50 based on U-Net architecture) to UAV orthophotos to automatically map floating plastic debris, achieving F1-scores of 0.86-0.92 for specific plastic types including oriented polystyrene, nylon, and PET. Classification accuracy decreased with lower spatial resolution, with 4 mm resolution providing optimal performance for distinguishing plastic types.