Sample laser speckle images generated from commercial milk in thin cuvette

This is a sample of the whole data used in the study "Combining transmission speckle photography and convolutional neural network for determination of fat content in cow milk - an exercise in classification of parameters of a complex suspension". Commercial milk samples of different fat content classes were purchased from supermarkets (where fresh milk was re-frigerated but ultra-pasteurized (UHT) milk was left on the open shelf). The samples were mainly from 3 dairy plants. We label the samples as Dairy 1 to Dairy 5 and a Collection of Dairies (CoD, see Table 1 in the paper). Dairies 4 and 5 were added to supplement the lack of 1.5% class in Dairies 1 and 2. The undiluted milk sample (10 μL) was loaded into a flat thin cuvette and covered with a 0.19 mm-thick coverslip. The cuvette was prepared by sticking a 50 μm-thick adhesive tape on a microscope slide and cutting out a 10 mm circular segment from the center of the tape. The sample in the cuvette was illuminated perpendicularly with a 5 mW, 1.7 mm collimated beam from a green frequency-doubled Nd:YAG laser at 532 nm. A 14-bit colour camera (Pike F-032C, AVT) was used to record laser speckle videos at 26.9o angle and 85 mm distance from the sample and at a constant temperature of 21±1oC. Two experimental runs were performed for each 1L-container of milk opened. In the first experimental run, the milk was loaded into the cuvette and 10 movies were recorded at 10 different locations on the sample (different background/static speckle) and labelled as “Version 1”. After recording 10 movies for Version 1 the cuvette was cleaned. It involved gently dipping it in a diluted cleaning solution, sonication for a minute and rinsing off with distilled water. Finally it was dried clean by blowing dry nitrogen gas (99.8%) over it. It was then loaded again with milk to record the movies for Version 2. At a single illumination point, ~1000 frames (13 s @ 75 fps) movie was recorded, the 10-locations on the same sample was exhaustive. Using milk form Dairy 1 only we experimented on 3 exposure time Protocols and confirmed our finding with milk from Dairies 2 and 3. The speckle images were extracted and normalized to grayscale level (0 - 1). Convolutional Neural Network (CNN) was trained on the version 1 of the data and tested on test set from version 1 and independent set (version 2). The study revealed that when "fully-developed laser speckles" are recorded from changing concentration and different particle sizes of suspensions CNN can be used to classify them unambiguously. In the folders there are sample images, videos, trained convolutional neural networks and classification algorithm

本数据集系对研究“结合透射散斑摄影与卷积神经网络测定牛奶中脂肪含量——复杂悬浮液参数分类之实践”所使用的全部数据的抽样。本研究所采用的商业牛奶样本来自不同脂肪含量级别的超市（其中新鲜牛奶需冷藏，而超高温巴氏杀菌（UHT）牛奶则可置于开放式货架）。样本主要源自三家乳制品厂。我们将样本标记为乳制品厂1至5，以及乳制品厂集合（CoD，详见论文中的表1）。乳制品厂4和5的加入旨在补充乳制品厂1和2中1.5%脂肪含量类别的不足。未经稀释的牛奶样本（10 μL）被装入扁平薄型比色杯中，并用0.19 mm厚的盖玻片覆盖。比色杯的制备方法为：将一张50 μm厚的粘合剂胶带粘贴在载玻片上，并从中裁剪出一个中心10 mm的圆形段。比色杯中的样本以垂直于样本的方式接受来自5 mW、1.7 mm准直的光束照射，该光束源自波长为532 nm的绿色频率倍增的Nd:YAG激光。使用14位彩色相机（Pike F-032C，AVT）在26.9°角度和85 mm距离处记录激光散斑视频，并在21±1°C的恒定温度下进行。针对每个开启的1L牛奶容器，进行两次实验运行。在第一次实验运行中，将牛奶装入比色杯，并在样本的10个不同位置（不同的背景/静态散斑）记录了10个电影，标记为“版本1”。在记录完版本1的10部电影后，对比色杯进行清洗，清洗过程包括将比色杯轻轻浸入稀释的清洗溶液中，超声处理一分钟，然后用蒸馏水冲洗干净。最后，通过吹入99.8%的干燥氮气将其彻底干燥。之后，再次装入牛奶以记录版本2的电影。在单一照射点，记录了约1000帧（13秒 @ 75 fps）的电影，对同一样本的10个位置进行了详尽的记录。仅使用乳制品厂1的牛奶，我们进行了三种曝光时间协议的实验，并使用乳制品厂2和3的牛奶验证了我们的发现。散斑图像被提取并归一化到灰度级别（0 - 1）。在大语言模型（LLM）上训练了数据版本1的卷积神经网络（CNN），并在版本1的测试集以及独立集（版本2）上进行了测试。研究结果表明，当从不同浓度和不同粒子大小的悬浮液中记录“完全发展的激光散斑”时，卷积神经网络（CNN）可以用来对这些散斑进行明确的分类。文件夹中包含样本图像、视频、训练好的卷积神经网络和分类算法。