Enhancing the Tracking of Seedling Growth Using RGB-Depth Fusion and Deep Learning

The use of high-throughput phenotyping with imaging and machine learning to monitor seedling growth is a tough yet intriguing subject in plant research. This has been recently addressed with low-cost RGB imaging sensors and deep learning during day time. RGB-Depth imaging devices are also accessible at low-cost and this opens opportunities to extend the monitoring of seedling during days and nights. In this article, we investigate the added value to fuse RGB imaging with depth imaging for this task of seedling growth stage monitoring. We propose a deep learning architecture along with RGB-Depth fusion to categorize the three first stages of seedling growth. Results show an average performance improvement of 5% correct recognition rate by comparison with the sole use of RGB images during the day. The best performances are obtained with the early fusion of RGB and Depth. Also, Depth is shown to enable the detection of growth stage in the absence of the light.

Trends in the development of sensor devices based on surface plasmon resonance (Review)

Trends in the development of modern sensory devices based on surface plasmon resonance (SPR) are considered. The basic principles of construction of SPR sensor are given. For excitation of surface plasmons on the surface of sensitive elements of biosensory, a prism of total internal reflection is used or a dielectric substrate are used. A thin (dozens nm) film of high-conductive metal (mainly gold or silver) is applied to the working surface of the prisms or dielectric substrate. In a typical observation experiment, SPR is measured dependence on the angle of increasing light intensity, reflected by the resonance sensitive surface of the prism (chip). The optical schemes and principles of work of various SPR sensors are considered: - SPR Sensors with angular modulation, which are the most commonly used method based on the corner registration, in which the SPR occurs. The surface of the metal film is irradiated by monochromatic light and scans on a certain range of angles. There is a kind of SPR sensors with angular modulation, in which there is no mechanical scan of the angle of fall. Such sensors are entirely necessary for excitation of PPRs a set of angles is obtained due to a divergent or convergent light beam. - PPR sensors with a wavelength modulation is based on fixing an angle of falling light at a certain value and modulation of the wavelength of the incident light. Excitation of surface plasmons leads to a characteristic failure in the spectrum of reflected radiation. - Phase sensitive SPR sensors in which a change in the phase of the light wave associated with the surface plasma is measured on one corner of the fall and the wavelength of the light wave and is used as the output signal. - SPR imaging sensors in which the Technology of SPR imaging (SPRi) combines the sensitivity of the SPR with spatial image capabilities. The SPRI circuit uses as a fixed angle (as a rule, a slightly left angle of the SPR) and a fixed wavelength to measure changes in the reflection ability (Δ% R) that occur when the curve of the SPR is shifted due to the change in the refractive index above the surface of the sensor element. - SPR imaging sensors polarization contrast. In order to improve the quality of high-performance SPR imaging sensors in terms of sensitivity and resolution, the method of polarization contrast is used Disadvantages and advantages of SPR sensors are constructed with different principles are considered. The design and prospect of the use of achromatic and suburchast wave plates in the PPR imaging sensors with polarization contrast are considered.

SPECTRAL MEASUREMENT AND CLASSIFICATION IN THE ERA OF BIG DATA

The measurement and classification of light is essential across many scientific disciplines. Devices used to measure light range from the highly precise scanning spectroradiometers to the more practical compact multichannel filter-array type imaging sensors and the ubiquitous RGB pixel. While there have been numerous successful efforts to reconstruct spectrum from RGB, RGB-to-spectrum reconstruction has historically been limited to natural scenes and other edge cases under strict constraints. However, information theory and recent advances in deep learning have shed new light on the vast amount of redundancy contained within data collected in the natural world, including light. In this paper, we will investigate how analytic methods can help map high dimensional spectra data to a low-dimensional feature space with minimal inductive bias. Through a better understanding of the intrinsic dimension of the data, we can use the features expressed in this representation to exploit regularities and make tasks like data compression, measurement and classification more efficient. The aim of this analysis is to help inform how and when low-dimensional representation of spectra is useful in practice for designing compact sensors as well as for lossy data compression and robust classification.

Synthetic Source Universal Domain Adaptation through Contrastive Learning

Universal domain adaptation (UDA) is a crucial research topic for efficient deep learning model training using data from various imaging sensors. However, its development is affected by unlabeled target data. Moreover, the nonexistence of prior knowledge of the source and target domain makes it more challenging for UDA to train models. I hypothesize that the degradation of trained models in the target domain is caused by the lack of direct training loss to improve the discriminative power of the target domain data. As a result, the target data adapted to the source representations is biased toward the source domain. I found that the degradation was more pronounced when I used synthetic data for the source domain and real data for the target domain. In this paper, I propose a UDA method with target domain contrastive learning. The proposed method enables models to leverage synthetic data for the source domain and train the discriminativeness of target features in an unsupervised manner. In addition, the target domain feature extraction network is shared with the source domain classification task, preventing unnecessary computational growth. Extensive experimental results on VisDa-2017 and MNIST to SVHN demonstrated that the proposed method significantly outperforms the baseline by 2.7% and 5.1%, respectively.

A Practical Validation of Uncooled Thermal Imagers for Small RPAS

Uncooled thermal imaging sensors in the LWIR (7.5 μm to 14 μm) have recently been developed for use with small RPAS. This study derives a new thermal imaging validation methodology via the use of a blackbody source (indoors) and real-world field conditions (outdoors). We have demonstrated this method with three popular LWIR cameras by DJI (Zenmuse XT-R, Zenmuse XT2 and, the M2EA) operated by three different popular DJI RPAS platforms (Matrice 600 Pro, M300 RTK and, the Mavic 2 Enterprise Advanced). Results from the blackbody work show that each camera has a highly linearized response (R2 > 0.99) in the temperature range 5–40 °C as well as a small (<2 °C) temperature bias that is less than the stated accuracy of the cameras. Field validation was accomplished by imaging vegetation and concrete targets (outdoors and at night), that were instrumented with surface temperature sensors. Environmental parameters (air temperature, humidity, pressure and, wind and gusting) were measured for several hours prior to imaging data collection and found to either not be a factor, or were constant, during the ~30 min data collection period. In-field results from imagery at five heights between 10 m and 50 m show absolute temperature retrievals of the concrete and two vegetation sites were within the specifications of the cameras. The methodology has been developed with consideration of active RPAS operational requirements.