DeepTrash

正浮力海洋塑料碎片的量化对于理解塑料是如何产生的至关重要
垃圾在世界海洋中堆积，对于识别目标热点也至关重要
清理工作。目前，量化海洋塑料最常用的方法是使用蝠鲼拖网
用于手动采样。然而，这种方法成本高，需要人工。这项研究
通过使用神经网络的自主方法消除了手动采样的需要，并且
计算机视觉模型，对从海柱各层捕获的图像进行训练
进行实时塑料量化。表现最好的模型具有平均平均精度
85% 和 0.89 的 F1 分数，同时保持接近实时的处理速度 2 ms/img。

Quantification of positively buoyant marine plastic debris is critical for understanding the origins of plastic pollution. The accumulation of waste in the world's oceans is also essential for identifying targeted marine pollution hotspots and supporting cleanup operations. Currently, the most widely adopted method for quantifying marine plastic involves manual sampling using a manta trawl. However, this approach is high-cost and labor-intensive. This study eliminates the need for manual sampling by employing an autonomous framework based on neural networks and computer vision models, which are trained on images collected from all layers of the water column to enable real-time plastic quantification. The top-performing model achieves a mean average precision (mAP) of 85% and an F1-score of 0.89, while maintaining a near-real-time processing speed of 2 milliseconds per image.

搜集汇总

背景与挑战

背景概述

DeepTrash是一个用于海洋塑料碎片识别的图像数据集，旨在通过计算机视觉模型实现实时塑料量化。该数据集由多个研究机构联合发布，最佳模型性能达到85%的平均精度和0.89的F1分数。