Open Access
Add to BibSonomy
Abstract
In the study, we have combined speckle metrology and deep learning tools in discriminating supermicroplastics (SMPs) sizes and concentrations. Polystyrene spheres used as SMPs were introduced in the container filled with salt water. The particles were illuminated with the 635 nm laser, and the scattered light was recorded with the CMOS camera. For the simulation studies, different sized particles (2 µm, 20 µm, and 200 µm) and concentrations were used. Speckles were analyzed using a deep learning algorithm to distinguish particles sizes and concentrations. It was demonstrated that the convolutional neural networks (CNNs) trained with speckles could distinguish feeble differences in speckle patterns depending on particle sizes and concentrations. Deep learning was found to be capable of distinguishing different particle sizes and concentrations from the speckle patterns. We suggest our combined technique could be effectively used in investigating MPs in the ocean where it remains challenging to conduct in situ surveys and obtain the SMP distribution in deeper regions of the ocean.
© 2022 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
Plastic products are ubiquitous in our life because they have a lot of advantages like lightweight, durability, and ease to mass-produce. In 2004, about 225 million tons of plastics were produced in the world, and the amount of production of plastics reached about 370 million tons in 2019 [1]. On the other hand, it was predicted that the amount of discarded plastic would also continue to increase [2]. If plastic products are discarded improperly, they enter the sea through rivers. While entering the sea and in the sea, they are degraded into fragments called microplastics (MPs) and supermicroplastics (SMPs) by UV rays, waves, and abrasion.
MPs are small plastic fragments whose size is equal to or less than 5 mm while SMPs are small plastic fragments whose size is equal to or less than 350 µm. Once MPs and SMPs are generated, it is difficult to collect them from the sea because they are small, translucent, or transparent. It means that MPs and SMPs are accumulated in the sea. It is estimated that the amount of plastic debris entering the sea would increase by an order of magnitude by 2025 compared to that in 2010 if waste management infrastructures in countries producing a large number of plastic debris are not improved [3]. Therefore, recently, marine pollution caused by MPs and SMPs is problematic in the world.
It is reported that when aquatic organisms like fishes are kept in the water tank that contains MPs, MPs were accumulated, and some adverse effects occurred, for example, short lifetime, inhabitation of growth, and low egg production [4,5]. Furthermore, MPs were identified from fishes and oysters that were taken from the sea [6,7]. These indicate that the effects of MPs occur not only on aquatic organisms but also on humans through the food chain. To investigate the effects of MPs and SMPs on organisms, it is necessary to survey the size of MPs & SMPs and the distribution of MPs & SMPs in the sea. However, there is not enough information on MPs & SMPs, especially SMP distribution in the sea due to technical challenges of the current survey of MPs and SMPs.
One of the current methods for surveying MPs in the sea is towing a net whose mesh size is around 350 µm followed by analyzing the collected samples by using microscopy, infrared spectrometer, and Raman spectrometer [8,9]. This method has mainly four following challenges. (1) It takes a lot of time to analyze the samples, (2) it is not able to collect SMPs, (3) it is not able to collect samples from the deep sea, (4) it has difficulty in separating organisms like algae from MPs and SMPs. Recently, the methods of detecting plastics using optics techniques in the water were demonstrated [10,11]. However, the methods for surveying SMPs have not been demonstrated yet.
In our previous work, we demonstrated that the speckle-based method could distinguish SMPs from creatures [12]. However, to survey SMPs using the speckle-based method, it is necessary to discriminate SMP size and concentrations from speckles. There exist only a few studies related to the application of speckles in microplastic classification. U-Net neural network has been used in quantifying microplastics of sizes 1 to 5mm from photographs [13]. Convolutional neural networks (CNNs) were applied to identify silicon dioxide and melamine particles of sizes of a few tens of micrometers from scattered light [14]. Yan et al. [15] presented a machine learning-based method to recognize suspension by distinguishing dispersoid-dependent speckle patterns produced by micron-sized protein powder and milk powder. In this study, an expanded work of [16], we have explored the potential of deep learning to discriminate SMP size and concentrations from speckles in salt water that simulate marine conditions.
2. Experiment
2.1 Experimental principle
When a rough object is illuminated with coherent light like a laser, light scatters and when the scattered light is observed at a given position far away from the object, a random granular pattern is observed. Such a random pattern is called the speckle. If the illuminated object is static, the intensity of the observed speckle is constant. However, when the illuminated object is in motion, then the intensity of the observed speckle is no longer a constant but fluctuates in accordance with the movement of the object.
The same thing is applied when tiny particles are illuminated with a laser. When tiny particles are illuminated with a laser, the light is scattered. If the scattered light is observed, speckles are observed. When tiny particles are in suspension-like water where they are in constant motion, the intensity of the observed speckle is dynamic or fluctuates.
In our previous work on distinguishing microorganisms from SMP [12], we could show that it is possible to discriminate from the difference in the speckle patterns depending on the SMP size added to the medium of the micro-organisms. Therefore, in this study, we explore the potential of deep learning as a tool to use speckle patterns in discriminating different SMPs depending on the sizes and volume fractions.
2.2 Experimental system
In this study, we carried out two experiments, a preliminary experiment and the main experiment. The preliminary experiment was carried out to test the possibility of applying deep learning as a survey method for SMP. The main experiment was carried out to investigate the extent of the limitations of the classification of SMP size and volume fraction by deep learning.
The experiment used for recording the speckles is shown in Fig. 1. A PMMA cuvette with a size of (mm): 12.5 × 12.5 × 45 having a capacity of 1.5~3.0 (mL) and optical pathlength of 10 mm and wavelength characteristics of 285–750nm containing the sample was placed in front of the laser of the wavelength of 635 nm of power of 15 mW attached with a variable lens (LDP6605-HA NEOARK Corporation, Tokyo, Japan). The lens was adjusted so that the laser beam was focused onto the surface of the cuvette and the beam size at the cuvette size was about 1 mm × 0.5 mm. The scattered light from the sample in the cuvette was recorded by a CMOS camera (1280 × 1024 pixels with a pixel size of 5.4 µm x 5.4 µm; CS135MU, Thorlabs Inc., NJ, USA). A variable iris was placed between the laser and the cuvette to prevent the CMOS camera from recording the stray light. The diameter of the variable iris was set to about 4 mm. Recordings were done six times. The first and fourth recordings were started 30 seconds after the cuvette was inverted gently and placed on the stand. Speckles were recorded as video data under 15 frames per second for 20 seconds. Recorded data were stored in a PC for analysis offline. The calculations here were done with a MacBook Pro (2019) that has a Quad Intel core i5, a processor with a speed of 1.4 GHz and a VRAM (approximately 1.54 GB).
Fig. 1. A schematic image of experimental system that was used for recording speckles of the main experiment.
Here, we would like to point out that a free space geometry was used for imaging speckles. Considering the beam size (a) at the cuvette (1 mm x 0.5 mm) and the distance (d) from the cuvette to the CMOS (280 mm), the approximate speckle size for 635 nm wavelength (λ) would be 110µm x 220 µm [0.61dλ/beam size ]. Considering the resolution of the camera pixels of 5.2 µm x 5.2 µm and using the full size of the camera 1024 × 1280 pixels with an image area of 6.66 mm(H) x 5.32 mm (V), speckle movies were recorded before resizing the images for analysis. As the speckles generated from the particles themselves would be much larger at the camera, making the system sensitive enough.
2.3 Sample preparation
It is expected that the size and concentrations of SMPs in the sea are different depending on the survey location. Therefore, we prepared nine samples that contained three different-sized polystyrene spheres (2 µm, 20 µm and 200 µm) and different volume fractions (0.01%, 0.001%, and 0.0001%). Volume fraction was used to adjust the volume of polyethylene spheres in each sample.
The samples used in the main experiment are shown in Table 1. Images of the samples are shown in supplementary Fig. 1. Polyethylene spheres used in this study were purchased ones (2 µm SMP and 20 µm SMP: Polysciences, Inc., PA, USA, catalog numbers are 19814-15 and 18329-5 respectively. 200 µm SMP: Thermo Fisher scientific. Inc., MA, USA, catalog number is 4320A). 2 µm and 20 µm SMPs were in a suspension form while 200 µm SMPs were of powder form.
Table 1. Samples used in the main experiment
Procedures for making samples were different depending on particle size and volume fraction due to the difference of the form of SMPs. Saltwater with a salinity of about 2% was used for making all samples to replicate the condition of the sea.
For the samples of 2µm SMP volume fraction 0.01% and 20 µm SMP volume fraction 0.01%, samples were made by following procedures. First, a volume of 2 µm SMP or 20 µm SMP calculated by Eq. (1) was put into the cuvette. Next, salt water was poured to fill the cuvette.
where Vc and Vs denote the volumes of the cuvette and that of the SMP respectively.For the sample of 200 µm SMP volume fraction 0.01%, first, about 5 mL of salt water was poured into a cuvette. Second, a volume of 200 µm SMP calculated by Eq. (1) was put into the cuvette. Then, the cuvette was filled with salt water.
For the rest of the samples, additional procedures were required because the pipette and the balance used in the experiment were not able to suck up the volumes of SMPs and measure the weight of SMPs required for making the samples respectively. For the samples of 2 µm SMP volume fraction 0.001% & 0.0001% and 20 µm SMP volume fraction 0.001% & 0.0001%, first, the samples volume fraction 0.01% were made. Second, 450 µL or 45 µL of liquid (450 µL for volume fraction 0.001% and 45 µL for volume fraction 0.0001%) was sucked up from the sample volume fraction 0.01% and was transferred to another empty cuvette. After that, the cuvette was filled with salt water.
For the samples of 200 µm SMP volume fraction 0.001% and 0.0001%, first, a sample volume fraction 0.01% was made. Second, a volume of liquid was removed from the cuvette so that a volume of liquid remained in the cuvette was 450 µL or 45 µL (450 µL for a volume fraction 0.001% and 45 µL for volume fraction 0.0001%). Then, the cuvette was filled with salt water. During recording, the following procedure was applied during the data collection:
- 1. Cuvette was gently manually inverted and put on the cuvette holder.
- 2. Waited about 30 sec.
- 3. Recorded the video sequentially 3 times. This was done to avoid any sampling errors during the recording itself.
- 4. The process was repeated starting from step 1.
2.4 Data set preparation
In order to make CNNs for classifying SMP size and volume fraction, transfer learning [16] was utilized. Training data sets and test data sets were made as follows.
2.4.1 Training data set preparation (preliminary experiment)
First, we made a training data set that included still images of three different sized SMPs (2 µm, 20 µm and 200 µm SMP) and salt water with salinity of about 2%.
The still images stored in the training data set were made by converting the video data that was recorded in our previous study [12] into still images. In the previous study, recordings of speckle video data were done six times following the same procedures as in the current study.
Figure 2 shows a schematic diagram of the training data sets. In the training data set, first 60 still images of each sample were used.
Fig. 2. A schematic diagram of the training data set obtained from earlier results [12] used in the preliminary experiment.
2.4.2 Training data set preparation (main experiment)
In the main experiment, two types of training data sets were made. Both training data sets consisted of ten kinds of speckle still images (samples shown in Table 1 and salt water with salinity of about 2%). However, the number of each kind of still image in the training data sets were varied.
Still images used in the training data sets were converted from the speckle video data that was recorded at the first recording. Figure 3 shows a schematic diagram of the training data sets.
In the first training data set (training data set A), the first 60 still images of each sample were stored. On the other hand, in the second training data set (training data set B), the first 100 still images of each sample were stored.
2.4.3 Test data set preparation (preliminary experiment and main experiment)
To investigate the classification accuracies of the trained CNNs, we made six test data sets. The test data sets for the preliminary experiment and the main experiment were composed of the same kind of sample still images as the training data sets. The still images in the test data sets of the preliminary experiment were converted from the video data that were recorded at the fourth and fifth recordings in the previous study [12]. The still images in the test data sets of the main experiment were converted from the video data that were recorded in this study at the fourth and fifth recordings. Figure 4 shows the speckle images recorded in the main experiment. The symbols used in Fig. 4 are written in Table 2. From Fig. 4 upon visual inspection, no large differences could be seen in the speckle patterns.
Table 2. Symbols used in Fig. 4
Figure 5 shows a schematic diagram of the test data sets of the preliminary experiment and the main experiment. The procedures for making test data sets for the preliminary experiment and the main experiment were the same as follows:
The first 100 still images, the second 100 still images, and the third 100 still images of the video data recorded at fourth recording were used for making the test data sets1, 2, and 3 respectively. The rest of the test data sets were made in the same way using video data recorded at the fifth recording.
3 Results and discussion
In this study, MATLAB (R2020b Update6) was used for training the pre-trained CNN and classifying the test data sets with a PC. In both experiments, training options for training the CNNs were the same.
Both in the preliminary and the main experiments GoogLeNet was used as the pre-trained CNN [17]. Out of the 22 layers, only two layers were modified, with the modifications done to the last output layer and its predecessor, the fully connected layer. Training of the CNNs was done until 30 epochs. When all the images stored are used for training the CNNs, it corresponds to an epoch. Thus, all the images in the training data set were used for training CNN over 30 times. The learning rate was set to be 0.001. The mini-batch size was set to 128. The stochastic gradient descent with momentum (SGDM) optimizer was used for the solver. Computation times involved are given in Appendix 1.
3.1 Preliminary experiment results
To investigate if there is a possibility that deep learning can be used as a tool for surveying SMP in the deep sea, first, we trained pre-trained CNN on the training data set that include three different sized SMP (2 µm, 20 µm, and 200 µm SMP) and salt water. Then, six test data sets that included three different-sized SMP (2 µm, 20 µm, and 200 µm SMP) and salt water were classified.
Figure 6 shows the classification results. Vertical labels and horizontal ones respectively are true classifications and predicted classifications. Values written in squares correspond to the mean number of classified speckle images. Each sample in the test data sets had 100 speckle images. Thus, the values written in the squares also mean the percentage of classified speckle images. Furthermore, the value written in the blue squares means classification accuracy. Squares without any values mean no speckle image was classified. From Fig. 6, all speckle images were correctly classified, and this result showed that deep learning could be utilized as a viable tool for surveying SMP.
Fig. 6. Classification result by the CNN trained with the speckles recorded in the preliminary experiment.
3.2 Main experiment results
Because of the positive results of high classification accuracy obtained in the preliminary experiment, in the main experiment, we made two training data sets (training data set A and training data set B). Both training data sets consisted of ten kinds of speckle images obtained for three different-sized SMPs, three-volume fractions, and saltwater. The difference between training data set A and training data set B was in the number of speckle images. In training data set, A, each sample had 60 speckle images, and in training data set B, each sample had 100 speckle images.
Figures 7 and 8 show respectively the classification results by the CNNs trained with the training data sets A and B. Vertical labels and horizontal ones respectively are true classifications and predicted classifications. Values written in squares are the mean number of classified speckle images. For each sample, the test data sets had 100 speckle images. Thus, the values written in the squares also mean the percentage of classified speckle images. Furthermore, the value written in blue squares means classification accuracy. Squares without any values mean non-classified speckle images. The symbols used in each figure are written in Table 3.
From Fig. 7, for the CNN trained with the training data set A, classification accuracies of all types of 2 µm SMP and 20 µm SMP were over 80%, high classification accuracy. From Fig. 7, for the trained CNN with the training data set B, classification results of all types of 2 µm SMP and 20 µm SMP except for 20 H were 100%.
Our classification results demonstrate that despite slightly different speckle patterns depending on SMP size and volume fraction, the trained CNNs might be able to recognize slight differences in the speckle patterns and thus distinguish particle sizes and volume fraction.
If the CNN trained with data set A could recognize slight differences in speckle patterns, then depending on SMP sizes and concentrations, we expect the features or the regions in the image that influence the classification would be different.
Here, we do see some discrepancies in the result with the misclassification of 200M to be misclassified as 200H and 200H to be misclassified as 20H and we consider two possible factors that could contribute to these discrepancies.
The sample preparation for 200 µm particles was done with powder and it is possible the sample itself is not completely homogenized as compared to 2 or 20 µm particles that came in liquid form and were thus completely homogenized. Even though we define the samples as high and low depending on the volume fractions, it is possible that within the illuminating volume there could be discrepancies with respect to the sample naming and the actual amount. Increasing the sample size could increase the sensitivity of the technique and of course, that would come with increased computation capabilities. Nevertheless, our study suggests the existing deep learning techniques could be successfully implemented in the discrimination of supermicroplastics.
The other factor that could possibly contribute to these discrepancies is that of comparing the scattering coefficients of the three particles used (Refer to Table 2 Ref. [12]). The scattering coefficient for 200 µm particles at the wavelength we use is almost two orders of magnitude lower compared to that for 2 or 20 µm SMPs. Therefore, the speckles generated would be affected greatly due to this difference as there will be less scattering from 200 µm particles. Therefore, depending on the position of the probing volume of the cuvette, there are chances for the existence of a larger number of particles or a smaller number of particles within the volume. This in turn might lead to the discrepancies that we saw in our results. As the particle sizes get larger, it may be necessary to make careful investigation with the algorithm and not just rely on the algorithm for classification. Other tools such as scattering measurements of size estimation may be necessary. Nevertheless, the method proved to be useful and sensitive even to a single particle within the probing region.
To visualize the features or regions in the images that influenced the classification, occlusion sensitivity maps were made. The occlusion sensitivity maps were made using a MATLAB function, occlusionSensitivity. Input arguments of occlusionSensitivity are a trained CNN, a speckle image to be classified, and the true label of the input speckle image. The CNNs trained with speckle images accept images of 224 × 224 pixels. Therefore, speckle images were resized to 224 × 224 pixels using MATLAB function, imresize before applying the occlusionSensitivity to the images.
Figure 9 shows the speckle images overlaid with occlusion sensitivity map of the CNN trained with data set A. Red parts of the images correspond to areas that were mainly used for the classification of SMP. On the other hand, blue parts correspond to areas that were not strongly used for the classification. As from Fig. 9, the position and size of red parts of 2 µm SMP and 20 µm SMP are different depending on SMP size and volume fraction. These results demonstrated that the CNN trained with speckle images could recognize slight differences in speckle patterns.
Compared to the classification results of 20H obtained by the training data sets A and B respectively in the Figs. 7 and 8, the classification accuracy decreased by about 10% for the case of B. This reduction could be considered as due to the lost or reduced features in the speckles images in comparison to set A. Such reduction could be attributed to the fluctuations in speckle intensity from the movement of SMPs and, possibly the speckle patterns that were slightly different despite the sample being the same. As a result, in the training data set B, the number of speckle images almost accounted for a larger percentage compared to the data set A. The same reason can be considered for the decrease in 200H classification accuracy in the result classified by the training data set B.
From Figs. 7 and 8, classification accuracies of 200L classified by the CNN trained on the data set A were low. However, the classification accuracy of 200L classified by the CNN trained on training data set B was of high accuracy with 83%. This may be caused by the number of SMP in the cuvette. Table 4 shows the number of SMPs for each of the samples used. As can be seen from the table, for the 200L, the number of particles is almost one. Therefore, even for the sample containing just a single particle, CNN could be effective depending on the training set used.
Table 4. The number of SMPs in each cuvette
As from Table 4 the number of 200 µm SMP is quite low compared to that of 2 µm SMP and 20µm SMP. Thus, it can be considered that speckle images of 200L in data set A were recorded while SMP did not pass over the laser. On the other hand, 40 speckle images of 200L stored in only the data set B were recorded while SMP passed. As a result, the classification accuracy of 200 L in Fig. 9 became 83%.
If the CNN trained with data set B could recognize slight differences in speckle patterns, parts that affected the classification should be also different depending on SMP sizes and concentrations. To visualize the parts that influenced the classification, occlusion sensitivity maps were made. Figure 10 shows the speckle images overlaid with the occlusion sensitivity map of the CNN trained with data set B.
As from Fig. 10, the position and size of red parts are also different depending on SMP size and volume fraction. This result shows that the CNN trained with data set B also recognized slight differences in speckle patterns despite it being difficult to distinguish speckle patterns visually.
4 Conclusions
In this study, two experiments were carried out to explore the potential of deep learning to discriminate SMP size and volume fraction from speckles and the following results were obtained:
- (1) The speckles recorded under different sizes of particles and concentrations were difficult to be discriminated visually. Despite that, our results demonstrate that trained CNN could recognize the slight differences in speckle patterns produced by particles of different sizes or concentrations implying that deep learning could indeed discriminate SMP size and volume fraction from speckles.
- (2) It was effective to analyze the speckles using deep learning even though there was a single particle in the cuvette. For the sample of 200L, it was estimated that there was a single particle in the cuvette and the classification result of 200 L classified by the CNN trained with data set B was about 80%.
In future, we will explore the potential of deep learning tools in the real-world classification capability with samples collected from field samples and extending the application for off-line analysis of data collected with submersible spheres proposed in our earlier work [12].
Appendix 1. Table showing the training time and classification time under different experimental conditions
| Experiment | Training timing | Classification time |
|---|---|---|
| Preliminary | 8’44” | 13” |
| Main: Dataset A | 31’ 59” | 30.3” |
| Main: Dataset B | 54’ 32” | 30.9” |
Funding
Environmental Restoration and Conservation Agency (JPMEERF21356447); Japan Society for the Promotion of Science (JP21H0155).
Acknowledgements
The article is an extension of the conference proceeding https://doi.org/10.1117/12.2609366
Disclosures
The authors disclose no potential conflicts of interest associated with this manuscript.
Data availability
Data underlying the results presented in this paper are not publicly available at this time but may be obtained from the authors upon reasonable request.
References
1. PLASTICS EUROPE, https://plasticseurope.org/knowledge-hub/plastics-the-facts-2015/, https://plasticseurope.org/knowledge-hub/plastics-the-facts-2020
2. R. Geyer, J. R. Jambeck, and K. L. Law, “Production, use, and fate to all plastics ever made,” Sci. Adv. 3(7), e1700782 (2017). [CrossRef]
3. J. R. Jambeck, R. Geyer, C. Wilcox, T. R. Siegler, M. Perryman, A. Andrady, R. Narayan, and K. L. Law, “Marine pollution. Plastic waste inputs from land into the ocean,” Science 347(6223), 768 (2015). [CrossRef]
4. Y. Cong, F. Jin, M. Tian, J. Wang, H. Shi, Y. Wang, and J. Mu, “Ingestion, egestion and post-exposure effects of polystyrene microspheres on marine medaka (Oryzias melastigma),” Chemosphere 228, 93–100 (2019). [CrossRef]
5. I. Varó, A. Perini, A. Torreblanca, Y. Garcia, E. Bergami, M. L. Vannuccini, and I. Corsi, “Time-dependent effects of polystyrene nanoparticles in brine shrimp Artemia franciscana at physiological, biochemical and molecular levels,” Sci. Total Environ. 675(675), 570–580 (2019). [CrossRef]
6. I. L. N. Bråte, D. P. Eidsvoll, C. C. Steindal, and K. V. Thomas, “Plastic ingestion by Atlantic cod (Gadus morhua) from the Norwegian coast,” Mar. Pollut. Bull. 112(1-2), 105–110 (2016). [CrossRef]
7. J. Patterson, K. I. Jeyasanta, N. Sathish, A. M. Booth, and J. K. P. Edward, “Profiling microplastics in the Indian edible oyster, Magallana bilineata collected from the Tuticorin coast, Gulf of Mannar, Southeastern India,” Sci. Total Environ. 691, 727–735 (2019). [CrossRef]
8. V. Hidalgo-Ruz, L. Gutow, and R. C. M. Thompson, “Thiel Microplastics in the marine environment: a review of the methods used for identification and quantification,” Environ. Sci. Technol. 46(6), 3060–3075 (2012). [CrossRef]
9. X. Wang, N. Bolan, D. C. W. Tsang, B. Sarkar, L. Bradney, and Y. Li, “A review of microplastics aggregation in aquatic environment: Influence factors, analytical methods, and environmental implications,” J. Hazard. Mater. 402, 123496 (2021). [CrossRef]
10. B. O. Asamoah, B. Kanyathare, M. Roussey, and K. E. Peiponen, “A prototype of a portable optical sensor for the detection of transparent and translucent microplastics in freshwater,” Chemosphere 231, 161–167 (2019). [CrossRef]
11. K. E. Peiponen, J. Räty, U. Ishaq, S. Pélisset, and R. Ali, “Outlook on optical identification of micro- and nanoplastics in aquatic environments,” Chemosphere 214, 424–429 (2019). [CrossRef]
12. D. Endo, T. Kono, Y. Koike, J. Yamada, and U. M. Rajagopalan, “Laser speckle imaging in discrimination of zooplanktons from supermicroplastics.,” Environ. Nanotechnol., Monit. Manage. 16, 100587 (2021). [CrossRef]
13. J. Grant-Jacob, Y. Xie, S. Benita, M. Praeger, M. D. T. McDonnell, D. J. Heath, M. Loxham, R. W. Eason, and B. Mills, “Particle and salinity sensing for the marine environment via deep learning using a Raspberry Pi,” Environ. Res. Commun. 1(3), 035001 (2019). [CrossRef]
14. G. Lee and K. Jhang, “Neural Network Analysis for Microplastic Segmentation,” Sensors 21(21), 7030 (2021). [CrossRef]
15. J. Yan, M. Jin, Z. Xu, L. Cheni, Z. Zhu, and H. Zhang, “Recognition Of Suspension Liquid Based On Speckle Patterns Using Deep Learning,” IEEE Photonics J. 13(1), 1–7 (2021). [CrossRef]
16. D. Endo, T. Kono, Y. Koike, H. Kadono, J. Yamada, and U.M. Rajagopalan, “Application of laser speckles and deep learning in discrimination of supermicroplastics and zooplanktons,” Proc. SPIE 11950, 1195007 (2022). [CrossRef]
17. MathWorks, https://www.mathworks.com/discovery/transfer-learning.html
