Regressing Image Sub-Population Distributions with Deep Learning.

Regressing the distribution of different sub-populations from a batch of images with learning algorithms is not a trivial task, as models tend to make errors that are unequally distributed across the different sub-populations. Obviously, the baseline is forming a histogram from the batch after having characterized each image independently. However, we show that this approach can be strongly improved by making the model aware of the ultimate task thanks to a density loss for both sub-populations related to classes (on three public datasets of image classification) and sub-populations related to size (on two public datasets of object detection in image). For example, class distribution was improved two-fold on the EUROSAT dataset and size distribution was improved by 10% on the PASCAL VOC dataset with both RESNET and VGG backbones. The code is released in the GitHub archive at achanhon/AdversarialModel/tree/master/proportion.

Imaging Aluminum Particles in Solid-Propellant Flames Using 5 kHz LIF of Al Atoms.

Laser-induced fluorescence imaging of aluminum atoms (Al-PLIF) is used to analyze the spatio-temporal behavior of aluminized solid propellant combustion. Using alternating LIF and chemiluminescence emission images of the particles in the gaseous and liquid phase evolving close to and far above the dynamically varying propellant surface, sequences of images were recorded and analyzed. The good sensitivity achieved enabled us to track the dynamics of the flame in the vicinity of particles detected all along the flame extension and up to 1.5 MPa. Analysis of wide-field images enabled droplet velocity measurements due to the high LIF sampling rate (5 kHz). The observed typical plume structures were in good agreement with alumina-formation prediction and previous shadowgraphy visualization. High-resolution sequences of images showed gaseous distribution behavior around the molten particles. The Al vapor phase was thus found to extend between 3 and 6.5 radii around the particles. Particle detachment dynamics were captured just above the propellant surface.

Quantitative field measurement of soot emission from a large gas flare using sky-LOSA.

Particulate matter emissions from unconfined sources such as gas flares are extremely difficult to quantify, yet there is a significant need for this measurement capability due to the prevalence and magnitude of gas flaring worldwide. Current estimates for soot emissions from flares are rarely, if ever, based on any form of direct data. A newly developed method to quantify the mass emission rate of soot from flares is demonstrated on a large-scale flare at a gas plant in Uzbekistan, in what is believed to be the first in situ quantitative measurement of soot emission rate from a gas flare under field conditions. The technique, named sky-LOSA, is based on line-of-sight attenuation of skylight through a flare plume coupled with image correlation velocimetry. Monochromatic plume transmissivities were measured using a thermoelectrically cooled scientific-grade CCD camera. Plume velocities were separately calculated using image correlation velocimetry on high-speed movie data. For the flare considered, the mean soot emission rate was determined to be 2.0 g/s at a calculated uncertainty of 33%. This emission rate is approximately equivalent to that of 500 buses driving continuously and equates to approximately 275 trillion particles per second. The environmental impact of large, visibly sooting flares can be quite significant.

Sky-scattered solar radiation based plume transmissivity measurement to quantify soot emissions from flares.

For gas flares typical of the upstream energy industry and similar point sources, most current methods for characterizing soot emissions are based on plume opacity rather than a quantitative measure of mass flux. The absence of more quantitative approaches is indicative of the inherent complexity of soot and the difficulties in characterizing emissions in an unbounded plume. A new experimental approach has been developed for the investigation of soot emissions in industrial plumes. Referred to as sky-LOSA, the diagnostic permits evaluation of 2D spatially resolved monochromatic sky-light transmissivity data over the width of a plume, where sky-light intensities behind the plume are obtained via an interpolation algorithm. By using Rayleigh-Debye-Gans Fractal Aggregate theory to relate transmissivity data to soot concentrations, and with knowledge of the velocity of the plume, it is possible to quantify mass flow rates of soot in a plume. Experiments on an unconfined lab-scale soot plume were used to support a detailed uncertainty analysis under a wide range of conditions and to estimate sensitivity limits of the technique. Results suggest field measurements of soot emission from flare plumes should be possible with overall uncertainties of less than 32%. This represents a significant advancement over existing techniques based on opacity measurements.