Open Access
Add to BibSonomy
Abstract
We explore an active illumination approach for remote and obscured material recognition, based on quantum parametric mode sorting and single-photon detection. By raster scanning a segment of material, we capture the relationships between each mirror position’s peak count and location. These features allow for a robust measurement of a material’s relative reflectance and surface texture. Through inputting these identifiers into machine learning algorithms, a high accuracy of 99% material recognition can be achieved, even maintaining up to 89.17% accuracy when materials are occluded by a lossy and multi-scattering obscurant of up to 15.2 round-trip optical depth.
© 2023 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
The field of three-dimensional (3-D) active imaging has made significant progression, being able to reconstruct physical scenes with immaculate precision. This development is required for the oncoming autonomous era as a means of artificial intelligence (AI) interacting with our world. However, spatial sensitivity alone is insufficient in determining the behaviour towards an object. For example, an autonomous car must know the difference between a pedestrian and a tree, a robotic arm must handle glass with care, and a security system must register a concealed weapon a threat while ignoring mundane objects [1,2]. At the moment, the majority of machine vision utilizes cameras and complex neural networks to identify objects within pictures. These techniques detect features such as shape, gradient, color, and reflectivity to classify objects [3–7]. However, they depend on bright illumination and condense data onto a two-dimensional plane, which limits spatial information [8]. Additionally, current camera-based material recognition systems fail to work through scattering and high-loss environments.
Commercially, the precision of laser metrology has been successful in identifying defects in manufacturing; however, they are often restricted to these environments due to their high power and eye safety restriction [9–11]. This defect detection is desirable in everyday life, especially interactions with people. The minute changes in roughness of a person’s skin detected through these techniques can indicate the difference between benign skin conditions such as seborrheic keratosis or potentially fatal melanomas, as well as general health evaluation [12–16]. Material recognition through optical emission mechanisms can be typically divided into back-scattered intensity analysis or spatial imaging distinctions, the later usually measured in metrics such as the root mean square (RMS) value of the surface deviations [12–15,17–19].
Light detection and ranging (LiDAR) systems innately capture both sets of data and have recently gained recognition for their ability to distinguish materials [1,19–21]. Recent single photon detection techniques create a reliable method of detecting individual photons and allow for active imaging despite poor flux of backscattered signal photons, with detection even ranging from tens of kilometers away [22–25]. However, these systems are often plagued by background noise and interference. The application of quantum parametric mode sorting (QPMS) is able to mitigate such signal to noise issues [26,27]. QPMS is a quantum frequency conversion at the phase matching edge, in which the signal distillation and wavelength transduction are all performed during a single pass of the picoseconds optical pulse through a nonlinear waveguide. In the past, it has demonstrated photonics signal detection with signal-to-noise ratio of 40 dB over a linear-optical filtering and detection system, surpassing the theoretical limit of an ideal matched linear optical filter by 11 dB [26,28].
Our past work has focused on refining this detection method and utilizing its strengths to various applications [19,21,25,26,28,29]. In particular interest to this work, we recently published a prior paper on material recognition through this technique [19]. That paper solely identified materials based on their time of flight (ToF) histogram information of a single pixel, which made it entirely dependent on porous surfaces with unique internal structures. This paper seeks to push past that limitation, primarily utilizing multi-pixel imaging and differing means of feature detection for material recognition.
2. Methods
Our methodology can divided into three primary components; the physical hardware of our system, our data collection process, and the data post-processing.
2.1 System specifications
Two trains of nearly transform-limited, 6 ps pulses at 1554.1 nm (probe) and 1565.5 nm (pump) are generated from a 50 MHz femtosecond mode-locked laser (Mendocino, Calmar laser) by using 200 GHz dense-wavelength-division-multiplexing (DWDM) filters. We measure the pulses’ intensity and phase profile in both spectral and temporal domains using a frequency resolved optical gating (FROG) pulse analyzer to calculate the mode selectivity of upconversion detection [26,28]. The collimated signal probe pulses at 1554.1 nm in a Gaussian beam of 2.2 mm diameter are transmitted to the target through a transceiver. A fiber circulator separates the outgoing signal pulses from the backscattered photons with a minimum isolation extinction of 55 dB. The transceiver is a simple monostatic coaxial set up utilizing off-the-shelf telecom-grade optical components. The backscattered photons are recombined with pump pulses through another DWDM and fiber-coupled into the mode-selective upconversion detector. A field-programmable-gate-array (FPGA) is used as a central processor for obtaining the data from the upconversion single photon detector (USPD). A complete schematic of the experimental setup can be viewed in Fig. 1. A more detailed list of these components is available in the Table S1 in Supplement 1.
Fig. 1. A diagram of the physical system, following the path as described in the previous paragraph. Additional acronyms used: MML, mode-locked fiber laser; WDM, wavelength-division-mutliplexing; ODL, optical delay line; EDFA, erbium-doped fiber amplifier; Si-APD, silicon avalanche photodiode; PPLN, periodically-poled lithium niobate.
2.2 Data collection
Four materials were selected based on their common everyday use individually as well as being the basis of more refined products; these materials being brick, glass, metal, and wood. Each material of interest was placed approximately 1.5 feet from the transceiver, in the center of the temporal window allotted by our optical delay. This delay ensures us an 18cm scanning length, where an automated program would identify its distance to millimeter accuracy and begin to roster scan the surface. It would record a square pattern of 29 by 29 pixels (6.5cm by 6.5cm), with each pixel varying in increments of 0.28 degrees (2.2mm) from the center. This raster pattern proceeded to scan the 20 picoseconds (6 mm) before and after the material’s center distance in single picosecond (0.3 mm) increments for comprehensive depth information. The pixel path with the highest photon counts was selected for a more extensive ToF measurement of 100 ps, with 50ps before and after the previously detected maximum peak. The data was collected in this specific manner to maximize collection in minimum run time; however, the processing is able to run on previous 3D scans not intended for material recognition as it is reliant on utilizing the QPMS potential through post processing. A raster scan for a single delay point took 12 seconds to complete, and each whole 100 ps scan needed approximately 20 minutes. By knowing the exact location of the object prior to scanning, a scan taking only the necessary data would take 4 minutes.
The materials were repositioned every other scan to ensure a new section of the surface was captured to generalize our data pool. Additionally the data collection process was several months long, allowing multiple parameters in both the system and materials to gradually change over time.
In order to demonstrate the true capabilities of the technique, the materials were then retested from behind a titanium dioxide (TiO2) block to serve as an obscurant; TiO2 being the pigment utilized in smoke grenades [30]. Both metal and glass were hidden behind a 7.6 optical depth (OD) block (15.2 OD round trip), while brick and wood were reduced to a 5 OD block (10 OD round trip) due to their naturally lower reflectivity. These obscurants were created in-lab, using TiO2 pigment and resin. The obscurant also broadens the beam by doubling its width, testing the retention of our system with differing spot sizes than initially trained. A view of our set up can be seen in Fig. 2. The count averages of each material can be seen in Table 1, with the obscured counterpart data in 2. We kept the average pixel counts of both cases relatively close to keep data consistent, this was done by increasing signal power in the obscured scenes.
Table 1. The average count values for the unobscured material measurements. The average counts are the average median pixel counts per scan, while the other is the average maximum pixel count per scan
Table 2. The average count values for the obscured material measurements. The average counts are the average median pixel counts per scan, while the other is the average maximum pixel count per scan
The success for material recognition in the unobscured and obscured was evaluated separately by training a different machine learning method for each one. For the unobscured case, a total of 800 scans were collected and randomly sorted into 600 training and 200 testing data. For the obscured case, all previous 800 unobscured scans were placed into training while the new 150 obscured scans were tested. The machine learning was only shown samples from the training case, then after learning it was tested on the previously unseen test case. Samples of each material’s ToF histogram under both conditions can be seen below in Fig. 3.
Fig. 3. Each subplot is an average ToF histogram of a material, unobscured and obscured. On average, the peak histogram full width at half maximum 8-10 ps, but after obscurant grows to 12-15ps.
2.3 Data processing
2.3.1 Time of flight for single pixel
A single pixel, one fixed mirror position, of each scan was selected to have its photon counting histogram used in entirety as raw data for material recognition. These scans were minimally processed, with each being normalized to its peak count. While we sampled a total 100 ps depth change, only a fraction is necessary as the full width at half maximum (FWHM) of the returning information is generally around 10 ps. The timing jitter of a typical time-correlated single photon counting system is limited by the timing jitter of the entire system influenced by the timing-jitter of the laser, detector and time-tagging electronics [26]. On the other hand, the time-jitter of the QPMS system is limited by pulse width of pump and signal as both of them are carved out from a single laser [25,26,28]. Note that in Fig. 3, the time-resolving upconversion photon counting reflects the intensity-correlation of pump and backscattered probe pulses, where the measured width is about 10ps, as opposed to approximately 50ps for typical Silicon APDs, is upconverted photons as a function of the temporal delay between the synchronous probe and pump pulses. It thus circumvents the limiting factors of detector time-jittering. The temporal features on ToF histogram are useful for material recognition, as the peak shape and its distortions provide insight into a permeable material’s internal structure [19]. However, for "hard" targets where light is reflected entirely at the surface, a single ToF photon distribution does not contain enough distinguishing features. Therefore, the later two traits consist of multiple ToF scans that are taken across the object’s surface, where information is extracted by their pixel by pixel changes.
The temporal feature on the ToF histogram will serve as a recognition factor, an example of our technique at its most fundamental level, as well as a comparison to measure the success of the matrix scanning.
2.3.2 Relative reflectivity
The highest photon count of each pixel’s photon histogram was recorded in a 29 by 29 matrix. As the size of our array directly limits the amount of information received, increasing or decreasing the number of pixels scanned would have the correlated effect on recognition success. Each matrix was individually normalized before converting to an image file where it is more easily understood by our machine learning methods, sample scans of which can be viewed in Fig. 4.
Fig. 4. Relative reflectivity surface scans: top left, brick; top right, glass; bottom left, metal; bottom right, wood. Visual distinctions can already be made prior to CNN algorithms. The color map is a normalized heat scale, where white is 1 and black is 0.
In homogeneous materials, changes in the peak photon count indicate a change in incident angle. The incident angle changes as we turn the mirror, with the highest counts correlating to a normal angle to the material’s surface. A truly flat object, such as the glass panel, results in a circular distribution decreasing in counts as our mirror angle increases. It is sensitive enough to even detect slight curvature in the metal sheeting used for the experiment. However, when the surface is rough, the surface normal is changing as well, resulting in a speckle field-like scattering intensity distribution when the incident beam and surface segment are normal to each other.
2.3.3 Surface roughness
The depth location of the highest photon count per pixel’s photon histogram is recorded in another 29 by 29 matrix. This results in a direct 3D reconstruction of the object’s surface, by assuming the most probable location of the object is the center of its photon peak distribution. Examples can be seen in the top of Fig. 5. Despite all targets having a flat surface, a natural curvature is seen in the rendering due to the increase in path length of the beam on angled routes, for which the photons require more time to return. This is fixed through the removal of gradual curvatures and results in an accurate analysis of the surface roughness where each pixel’s deviations can be compared against its peers, shown in the bottom of Fig. 5. The rougher a material’s surface, the more prominent and frequent the deviations. By taking the root-mean-square value of each of the curve-corrected pixels, we get a succinct quantitative description of each material’s roughness. Theses matrices were once again converted to images prior to being inserted into its respective machine learning method. However, in the obscured case the data underwent a median filter to reduce noise prior to image conversion.
Fig. 5. Top: A direct 3D reconstruction of the brick’s surface is plotted by utilizing the center location of each returning pixel’s ToF histogram. Bottom: gradual curvature from increased path length is removed. The color map is a normalized heat scale, where white is 1 and black is 0, a color metric is provided as well
It is important to note the limits of our depth sensitivity. Nonlinear optical temporal gating as well as our incremental step size place seemingly only allow for 10 picoseconds (3 mm) temporal resolution for each measurement. However, by sampling high repetition rate laser pulses and performing statistical averages to reduce the shot noise of the photon arrival time measurement, derivations of our system have been able to register displacements of 110 nm [29]. Even the slightest deviation of the target’s surface brings about a change in returning beam’s angle, which directly affects the number of photons counted. With our noise removal capabilities to ascertain the correct photons and sensitivity to count each one, these low photon flux alterations are accounted for, allowing our system to surpass its physical limitations seemingly restricted by the pulse width and step size.
2.3.4 Machine learning
The machine learning is modular in nature, consisting of 3 classifiers individually identifying the material based upon their given trait, consisting separately of either time of flight histograms, relative reflectivity, or surface roughness. They then output a percentage chance the test scan has of being a given material.
For the ToF classification, the selected pixel’s ToF histogram was inserted into a K Nearest Neighbor (KNN) algorithm. A KNN is a non-parametric supervised machine learning method for estimating the likelihood of a datapoint being a member of a cluster based on its distance to said clusters. There were four neighbors to reflect the four materials classified. For training, we used a uniform distance weight and all points in each neighborhood were equally weighted.
The later traits were classified through a convolutional neural network (CNN). The architecture of a CNN was selected for its specific success in material identification compared to similar networks [31]. We utilized a pre-built CNN structure directly from a MATLAB example with the only changes being to the input and output layers to match our experimental system of 29 by 29 pixel images with 4 output classes [32]. We specifically avoided tailoring the machine learning aspects, which would most likely strengthen classification success, in order to have the results be more dependent on our system’s innate ability to collect patterns from the materials.
For our CNN architecture, we have three convolution layers, two max pooling layers, and one fully connected layer leading to soft max output. Our input layer is one channel, consisting of the normalized photon counts now registered as pixel values in our 29 by 29 images. This is followed by a 2D convolutional layer with a 3 by 3 filter, where said filter is swept across the input data to compute 8 feature maps made of sub-matrices in the filter’s size. Afterwards we perform batch normalization, an operation that zero-centers and normalizes each feature map. We then use a rectified linear unit (ReLU), an activation function that introduces non linearity and combats vanishing gradients, a common issue with back propagation. We end our first cycle with a 2D max pooling layer, which performs down-sampling by dividing the normalized feature maps into pooling regions, then computes the maximum of each region. We repeat derivations of this process twice more with increasing feature maps (16 then 32). Finally, we create a fully connected layer with 4 neurons which allows us to classify our four materials. We compute classification values with softmax, a neural transfer function which converts a vector of N numbers into a probability distribution of N possible outcomes. MATLAB then uses an additional layer to return the classification with the highest probability. A visual of this architecture can be seen in Fig. 6.
The fundamental convolution layer of the convolution neural network means that the classification is not determined by individual pixel values, but the relationships between its neighbors. Each pixel’s value is compared to those around it, and patterns that form from the deviations become identifying characteristics.
We used stochastic gradient descent (SGD) as a model optimizer, to iteratively minimize loss during back-propagation, with a learning rate of 0.01. Furthermore, we train for 50 epochs and shuffle the data each epoch. Note that the required CNN’s architecture and training effort can be different depending on the experimental parameters such as optical pulse width, temporal resolution and efficiency of the detector, as well as the reflectivity of the target. Our efforts for this paper were to increase recognition by progressing our system physically, in lieu of making neural net optimizations.
3. Results
Our neural nets trained and tested using the unobscured data sets had an accuracy of 99%, 98.5%, and 87% for their given traits of relative reflectivity, surface roughness, and ToF respectively. The results for the obscured test data trained with unobscured data on these same networks were 89.17% for both the relative reflectivity and surface roughness traits, while the ToF accuracy was reduced to 57% accuracy. The confusion matrix for the relative reflectance can be seen in Fig. 7 and the confusion matrix for the surface roughness in Fig. 8. A confusion matrix is a useful tool for viewing the accuracy of correct classification as well as mistaken selections made by the program. The matrix compares the actual target materials with the predictions by the machine learning model. A perfect model would have the predicted and true class being the same. However, this matrix is most useful in imperfect models, where it visually demonstrates the inaccuracies of the model. There it demonstrates which targets are "confused" for one another and therefore can locally diagnose inaccuracies.
We used pre-built CNN structures so the recognition’s success depended on the strength of our captured data’s underlying patterns, rather than post-processing optimizations. By heavily obscuring the capturing process of the testing sample, the high accuracy demonstrates a robustness of these patterns despite only being trained on unobscured data. While the individual ToF of each pixel suffers through the distortions,the relationships between them are preserved.
We can understand the surface roughness matrices through RMS values for the non-transparent materials. The values in Table 3 depict the average RMS value of each material in both unobscured and obscured scenarios. The values are fairly consistent despite the noise and scattering from the obscurant. The relatively small standard deviations also show persistence between scans, with low chance of associated material values overlapping. RMS values from 29 by 29 depth matrices can be reduced to 3 by 3 segments, with only a 0.003 ps deviation from the whole RMS value. This could potentially lead to a 98.9% reduction in scanning time in future endeavors.
Table 3. The RMS values for each material with their standard deviation. Glass is excluded due to its unique transparency as well as dual surface features; the RMS calculation of glass inadvertently measures its thickness.
4. Discussion
The significant loss in individual ToF recognition is due to the scattering that the optical pulses undergo throughout the medium, leading to huge signal attenuation. Photons attenuated to levels below the dark count are then enshrouded. Additionally, the pulse width of the photon arrival time histogram broadens by approximately 50% after the obscurant due to scattering, which can bury any identifying distortions the object may give it. While the individual ToF identifications suffer, the relationships between the pixels are maintained. The optical signal’s power is significantly weakened due to scattering, but the diffusion of this probe beam is undergone in all pixels, and when re-normalized then resembles the unobscured group identity the neural net was trained on. The specific nature of a CNN allows for these patterns, which are maintained despite the detrimental environmental obstacles, to be found.
The broadening is not solely temporal, as the beam spot size also increases after the obscurant. The success in the face of both broadening effects demonstrates the robustness of the system trained under more refined pulse widths and spot sizes. This indicates the performance of the system is not wholly dependent on our pre-selected spot size or pulse duration.
As for the matrix scanning results, the materials of brick and wood were least accurate in the perspective of reflectance. This can be attributed to the materials’ low reflectance allowing the in-homogeneity of the TiO$_2$ block to influence the readings. In the roughness classifier, the CNN confused wood and metal as the scattering altered the depth information and their RMS values only differed by 0.3ps, as shown in Table 3. The glass, while smooth, has its 3D profile identify both the front and back of its thin frame, often alternating the most prominent peak location. This makes it look either smooth or rough depending on the case, which is a distinct trait not missed in its own identification but confusing for other material assessments. The RMS table also offers insight to slight mislabeling of brick obscured scans, as the greatest shift between unobscured and obscured cases appears in the brick surface structure. This is most likely due to the scattering effect having the greatest discrepancies over its non-uniform surface.
The utilization of a transceiver set up allows us to reduce the angles of interest to only the laser’s incident angle relative to the target’s surface. While we only varied 3.5$^{\circ }$ from normal in the relevant dimensions, this technique is not limited to these angle ranges. The single-photon sensitivity of the system allows the return of the relevant information as long as the target is not a perfect mirror. The greater the angle of incidence, the less photons returned, but this can be negated through larger dwell times or higher power. The field of view of our system is intrinsically limited by the effective aperture of the aspheric lens and numerical aperture (NA) of the SMF consisting of the transceiver.
It is important to note the differences between our system and that of speckle contrast imaging (SCI). In SCI, the interest is in the speckle pattern generated by a coherent beam of light of uniform incident angle hitting a rough surface, to which the returning wavelets interfere and generate a unique pattern registered by a multi-pixel camera. This pattern is strongly dependent on this unique incident angle of the beam. In our system, we only acquire the photon count of in a single spatial mode as defined by the optical lenses and single mode fiber in the receiver, a single pixel. Every mode corresponds to an unique incident angle whose photon count is recorded to cumulatively form the speckle-like image, but is ultimately 841 separate measurements with intra-pixel inference due to the relatively small beam size and no inter-pixel interference (as photons in different pixels are recorded at different time). While SCI records the fingerprint of a material for a given angle, we describe how the qualities change over varying ones. The robustness of our system is demonstrated throughout the experiment. Training only on unobscured data, we are still able to accurately identify the materials despite heavy attenuation and scattering that visually block the targets. This occluding broadens both the ToF histograms and the beam size, showing resilience even in a combination of these cases. The targets themselves were scanned at varying distances from the transceiver over the course of months which also allowed numerous parameters in both the system and materials to gradually change while the classification accuracy remained high.
5. Conclusion
In general, LiDAR technology is used for spatial reconstruction and distance measurements. However, the returning information can provide much more than what was initially sought after. Traits such as relative reflectivity and surface roughness can be acquired by consequence, and when applied with machine learning provide unique identifying features that are preserved even when visually obscured. This allows LIDAR material recognition to surpass a fundamental limitation of traditional camera-based computer vision which will always require a clear visual path of sight. In conclusion, we have demonstrated a novel method of object recognition that once registered in an ideal setting can be applied to unforeseen photon-starved situations, allowing one to faithfully recognize hidden targets using only a neural network that was trained in an unobscured setting. These results indicate a robust material recognition system capable of module recognition through two different inputs. Furthermore, a combination of faster scanning time, low power signal, and "eye-safe" wavelength paired with the ability to work in high loss environments remove the restrictions of high-power alternatives. These innate safety measures make it applicable to pedestrian situations. In particular, this technique has biomedical potential as surface recognition continues to become a diagnostic tool. Future work will see roughness measurements carried out in more detail and its qualitative potential explored.
Funding
U.S. Army Combat Capabilities Development Command (W15QKN-18-D-0040).
Acknowledgments
We thank Michelle Wang for assistance in early data collection. This material is based upon work supported by the ACC-New Jersey under Contract No. W15QKN-18-D-0040.
Disclosures
The authors declare no conflicts of interest.
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.
Supplemental document
See Supplement 1 for supporting content.
References
1. K. Kidono, T. Miyasaka, A. Watanabe, T. Naito, and J. Miura, “Pedestrian recognition using high-definition lidar,” 2011 IEEE Intelligent Vehicles Symposium (IV), (2011), pp. 405–410.
2. D. Guo, H. Liu, B. Fang, F. Sun, and W. Yang, “Visual affordance guided tactile material recognition for waste recycling,” IEEE Transactions on Automation Science and Engineering pp. 1–9 (2021).
3. S. Bell, P. Upchurch, N. Snavely, and K. Bala, “Material recognition in the wild with the materials in context database,” arXiv, arXiv:1412.0623, 2014, [CrossRef] .
4. A. Dimitrov and M. Golparvar-Fard, “Vision-based material recognition for automated monitoring of construction progress and generating building information modeling from unordered site image collections,” Adv. Eng. Informatics 28(1), 37–49 (2014). [CrossRef]
5. H. Zhang, K. Dana, and K. Nishino, “Reflectance hashing for material recognition,”, in 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), (2015), pp. 3071–3080.
6. J. Tou, P. Tou, P. Y. Lau, and Y. H. Tay, “Computer vision-based wood recognition system,” Proc. Int’l Workshop on Advanced Image Technology (2007).
7. J. Xue, H. Zhang, K. J. Dana, and K. Nishino, “Differential angular imaging for material recognition,” Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2016).
8. F. Martino, C. Patruno, N. Mosca, and E. Stella, “Material recognition by feature classification using time-of-flight camera,” J. Electron. Imaging 25(6), 061412 (2016). [CrossRef]
9. Y. Li, H. Yan, and Y. Zhang, “A deep learning method for material performance recognition in laser additive manufacturing,” 2019 IEEE 17th International Conference on Industrial Informatics (INDIN), vol. 1 (2019), pp. 1735–1740.
10. V. Rodríguez-Fajardo, C. Rosales-Guzmán, O. Mouane, D. Wamwangi, E. Sideras-Haddad, F. S. Roux, and A. Forbes, “All-digital 3-dimensional profilometry of nano-scaled surfaces with spatial light modulators,” Appl. Phys. B 127(11), 145 (2021). [CrossRef]
11. M. V. Svechnikov, N. I. Chkhalo, M. N. Toropov, N. N. Salashchenko, and M. V. Zorina, “Application of point diffraction interferometry for middle spatial frequency roughness detection,” Opt. Lett. 40(2), 159–162 (2015). [CrossRef]
12. L. Tchvialeva, W. Shenkenfelder, A. Doronin, I. Meglinski, I. Markhvida, and T. K. Lee, “In vivo laser speckle measurements of skin roughness: clinical study,”Optics in the Life Sciences, (Optical Society of America, 2015).
13. A. S. Lotay, M. T. Carvalho, and J. M. Girkin, “Non-invasive assessment of skin roughness through speckle pattern analysis,” Biomedical Optics 2016, (Optical Society of America, 2016).
14. M. Zieger, M. Kaatz, S. Springer, R. Riesenberg, A. Wuttig, M. Kanka, S. Stanca, C. Reble, G. Khazaka, R. Sieg, M. De Gregorio, M. Sattler, and F. Fischer, “Multi-wavelength, handheld laser speckle imaging for skin evaluation,” Skin Res Technol 27(4), 486–493 (2021). [CrossRef]
15. D. C. Louie, J. Phillips, L. Tchvialeva, S. Kalia, H. Lui, W. Wang, and T. K. Lee, “Degree of optical polarization as a tool for detecting melanoma: proof of principle,” J. Biomed. Opt. 23(12), 1–7 (2018). [CrossRef]
16. G. Dhadwal, “In vivo laser speckle measurement of skin roughness versus profilometric measurement of silicone replicas,” J. Am. Acad. Dermatol. 68, AB36 (2013).
17. A. Jagannathan, A. J. Gatesman, and R. H. Giles, “Characterization of roughness parameters of metallic surfaces using terahertz reflection spectra,” Opt. Lett. 34(13), 1927–1929 (2009). [CrossRef]
18. X. Li and Y. Liang, “Remote measurement of surface roughness, surface reflectance, and body reflectance with lidar,” Appl. Opt. 54(30), 8904–8912 (2015). [CrossRef]
19. D. Tafone, I. Huang, P. Rehain, S. Zhu, Y. M. Sua, and Y. Huang, “Single-point material recognition by quantum parametric mode sorting and photon counting,” Appl. Opt. 60(14), 4109–4112 (2021). [CrossRef]
20. C. R. Yu and C. C. Peng, “Experimental study of material classification and recognition by a lidar scanning,” 2020 IEEE International Conference on Consumer Electronics - Taiwan (ICCE-Taiwan), (2020), pp. 1–2.
21. Y. M. Sua, S. Zhu, P. Rehain, D. Tafone, B. Muthuswamy, J. Ramanathan, I. Dickson, and Y.-P. Huang, “Quantum 3D imaging through multiscattering media of 10 optical depth,” Proc. SPIE 11420, 1142009 (2020). [CrossRef]
22. D. Faccio and A. Velten, “A trillion frames per second: the techniques and applications of light-in-flight photography,” Rep. Prog. Phys. 81(10), 105901 (2018). [CrossRef]
23. D. Shin, F. Xu, D. Venkatraman, R. Lussana, F. Villa, F. Zappa, V. K. Goyal, F. N. C. Wong, and J. H. Shapiro, “Photon-efficient imaging with a single-photon camera,” Nat. Commun. 7(1), 12046 (2016). [CrossRef]
24. Z.-P. Li, X. Huang, Y. Cao, B. Wang, Y.-H. Li, W. Jin, C. Yu, J. Zhang, Q. Zhang, C.-Z. Peng, F. Xu, and J.-W. Pan, “Single-photon computational 3d imaging at 45 km,” Photonics Res. 8(9), 1532 (2020). [CrossRef]
25. P. Rehain, Y. M. Sua, S. Zhu, I. Dickson, B. Muthuswamy, J. Ramanathan, A. Shahverdi, and Y.-P. Huang, “Noise-tolerant single photon sensitive three-dimensional imager,” Nat. Commun. 11(1), 921 (2020). [CrossRef]
26. A. Shahverdi, Y. M. Sua, L. Tumeh, and Y.-P. Huang, “Quantum parametric mode sorting: Beating the time-frequency filtering,” Sci. Rep. 7(1), 6495 (2017). [CrossRef]
27. F. Hlawatsch and W. Kozek, “Time-frequency projection filters and time-frequency signal expansions,” IEEE Trans. Signal Process. 42(12), 3321–3334 (1994). [CrossRef]
28. A. Shahverdi, Y. M. Sua, I. Dickson, M. Garikapati, and Y.-P. Huang, “Mode selective up-conversion detection for lidar applications,” Opt. Express 26(12), 15914–15923 (2018). [CrossRef]
29. P. Rehain, J. Ramanathan, Y. M. Sua, S. Zhu, D. Tafone, and Y.-P. Huang, “Single-photon vibrometry,” Opt. Lett. 46(17), 4346–4349 (2021). [CrossRef]
30. B. G. DeLacy, D. Redding, and J. W. Matthews, “Optical, physical, and chemical properties of surface modified titanium dioxide powders,” DTIC Document (2011).
31. A. Shukla, G. Kalnoor, A. Kumar, N. Yuvaraj, R. Manikandan, and M. Ramkumar, “Improved recognition rate of different material category using convolutional neural networks,” Mater. Today: Proc. 81(2), 947–950 (2023). [CrossRef]
32. Mathworks, “Create simple deep learning network for classification,” https://www.mathworks.com/help/deeplearning/ug/create-simple-deep-learning-network-for-classification.html.
