DeepParticle dataset [MICRO, MESO, MACRO] (2022)

These are the datasets used to train the models described in our method paper: "Computer vision segmentation model—deep learning for categorizing microplastic debris" The image segmentation and classification models are designed to identify plastic particles from an image. Two types of models were built: Segmentation models which segment and label the particle found within an input image with a homogeneous backgroundClassification models that classify a particle image into a category (e.g. hard plastic, pellet, line, cap...)See our Gitlab page: https://gitlab.com/Grouumf/particle_detect Each folder: MICRO, MESO, and MACRO represents a different size category. MICRO: 0.05 cm–0.5 cmMESO: 0.5 cm–5.0 cmMACRO > 5.0 cmThe VALIDATION dataset contains additional annotated images used for validation. Each folder is organized as follows: MICRO ├── annotation ├── foam ├── hard ├── line ├── noise ├── pellet ├── raw_img └── reference `raw_img` contains the raw images that each display a variety of plastic particles.`annotation` contains the particle pixel coordinates within the raw images.The remaining folders represent particle categories (line, noise, pellet, foam, reference). Each of these folders contains the individual image of each particle annotated in the raw imagesFor example, the MICRO annotation folder contains a TSV file named `08_03_micro_line_only1.JPG.tsv`. This file contains the particle annotations and coordinates for the raw images labeled: `08_03_micro_line_only1.JPG` The beginning of the TSV file is as follows: ID minr minc maxr maxc class line/2022-03-28 07:18:39.577051_k0.jpg 2255 530 2329 601 line line/2022-03-28 07:18:39.577051_k1.jpg 2185 754 2363 842 line line/2022-03-28 07:18:39.577051_k2.jpg 2974 863 3079 921 line Each line represents an annotated particle, with the four first columns its pixel coordinates. the last column is the annotated category. Also, the image of the first annotated particle `line/2022-03-28 07:18:39.577051_k0.jpg` can be found in the `line` folder. The MESO and MACRO folders contain different categories, e.g. cap_me, or butt_me, which are the bottle caps and the cigarette butt used for the MESO models. Finally, the noise folders contain noise "particles" or sections of the raw images that should not be identified as particles. These are useful when training the classifiers and removing the false positives from the annotations.本数据集用于训练本方法论文中提及的模型：《计算机视觉分割模型——用于微塑料碎屑分类的深度学习》。 该图像分割与分类模型旨在从图像中识别塑料颗粒。本次共构建两类模型： 一类为分割模型，可对输入图像中的颗粒进行分割与标注，输入图像背景为均质背景；另一类为分类模型，可将单颗粒图像归类至对应类别（如硬质塑料、颗粒、条状物、瓶盖等）。 详见我们的Gitlab页面：https://gitlab.com/Grouumf/particle_detect。 MICRO、MESO和MACRO三个文件夹分别对应不同的尺寸类别： - MICRO：0.05 cm–0.5 cm - MESO：0.5 cm–5.0 cm - MACRO：>5.0 cm 验证集（VALIDATION）包含额外的标注图像，用于模型验证。 以MICRO文件夹为例，其内部组织结构如下： MICRO ├── annotation ├── foam ├── hard ├── line ├── noise ├── raw_img └── reference 其中，`raw_img`存放包含多种塑料颗粒的原始图像；`annotation`存放原始图像中颗粒的像素坐标信息；其余子文件夹对应不同的颗粒类别（line、noise、pellet、foam、reference），每个文件夹均存放从原始图像中标注得到的单颗粒图像。 以MICRO文件夹下的annotation子文件夹为例，其中包含一个名为`08_03_micro_line_only1.JPG.tsv`的TSV（制表符分隔值）文件，该文件存储了对应原始图像`08_03_micro_line_only1.JPG`的颗粒标注信息与像素坐标。该TSV文件的开头格式如下： ID minr minc maxr maxc class line/2022-03-28 07:18:39.577051_k0.jpg 2255 530 2329 601 line line/2022-03-28 07:18:39.577051_k1.jpg 2185 754 2363 842 line line/2022-03-28 07:18:39.577051_k2.jpg 2974 863 3079 921 line 每一行对应一个被标注的颗粒，前四列为该颗粒的像素坐标，最后一列为其标注类别。此外，首个被标注的颗粒图像`line/2022-03-28 07:18:39.577051_k0.jpg`可在`line`子文件夹中找到。 MESO与MACRO文件夹包含其他颗粒类别，例如cap_me与butt_me，分别对应用于MESO模型训练的瓶盖与烟头。 最后，noise文件夹存放被标记为噪声的“伪颗粒”或原始图像中的非颗粒区域，可用于分类模型的训练，并用于过滤标注结果中的假阳性样本。