MétaCan
Menu
Back to cohort
Record W3016082095

Green Function and Electromagnetic Potential for Computer Vision and Convolutional Neural Network Applications

2019· article· en· W3016082095 on OpenAlex

Why this work is in the frame

A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.

fundA Canadian funder is recorded on the work.
no affNo Canadian affiliation: this work is invisible to an affiliation-only frame.
No Canadian affiliation. An affiliation-only frame, the usual design, would never have seen this work. It is one of the works that make the case for inverting the frame.

Bibliographic record

VenuePolyPublie (École Polytechnique de Montréal) · 2019
Typearticle
Languageen
FieldEngineering
TopicInfrared Target Detection Methodologies
Canadian institutionsnot available
FundersFonds de recherche du Québec – Nature et technologiesNatural Sciences and Engineering Research Council of Canada
KeywordsHumanitiesPhysicsArtificial intelligenceComputer scienceArt
DOInot available

Abstract

fetched live from OpenAlex

RESUME Pour les problemes de vision machine (CV) avancees, tels que la classification, la segmentation de scenes et la detection d’objets salients, il est necessaire d’extraire le plus de caracteristiques possibles des images. Un des outils les plus utilises pour l’extraction de caracteristiques est l’utilisation d’un noyau de convolution, ou chacun des noyaux est specialise pour l’extraction d’une caracteristique donnee. Ceci a mene au developpement recent des reseaux de neurones convolutionnels (CNN) qui permet d’optimiser des milliers de noyaux a la fois, faisant du CNN la norme pour l’analyse d’images. Toutefois, une limitation importante du CNN est que les noyaux sont petits (generalement de taille 3x3 a 7x7), ce qui limite l’interaction longue-distance des caracteristiques. Une autre limitation est que la fusion des caracteristiques se fait par des additions ponderees et des operations de mise en commun (moyennes et maximums locaux). En effet, ces operations ne permettent pas de fusionner des caracteristiques du domaine spatial avec des caracteristiques puisque ces caracteristiques occupent des positions eloignees sur l’image. L’objectif de cette these est de developper des nouveaux noyaux de convolutions bases sur l’electromagnetisme (EM) et les fonctions de Green (GF) pour etre utilises dans des applications de vision machine (CV) et dans des reseaux de neurones convolutionnels (CNN). Ces nouveaux noyaux sont au moins aussi grands que l’image. Ils evitent donc plusieurs des limitations des CNN standards puisqu’ils permettent l’interaction longue-distance entre les pixels de limages. De plus, ils permettent de fusionner les caracteristiques du domaine spatial avec les caracteristiques du domaine du gradient. Aussi, etant donne tout champ vectoriel, les nouveaux noyaux permettent de trouver le champ vectoriel conservatif le plus rapproche du champ initial, ce qui signifie que le nouveau champ devient lisse, irrotationnel et conservatif (integrable par integrale curviligne). Pour repondre a cet objectif, nous avons d’abord developpe des noyaux convolutionnels symetriques et asymetriques bases sur les proprietes des EM et des GF et resultant en des noyaux qui sont invariants en resolution et en rotation. Ensuite, nous avons developpe la premiere methode qui permet de determiner la probabilite d’inclusion dans des contours partiels, permettant donc d’extrapoler des contours fins en des regions continues couvrant l’espace 2D. De plus, la presente these demontre que les noyaux bases sur les GF sont les solveurs optimaux du gradient et du Laplacien.----------ABSTRACT For advanced computer vision (CV) tasks such as classification, scene segmentation, and salient object detection, extracting features from images is mandatory. One of the most used tools for feature extraction is the convolutional kernel, with each kernel being specialized for specific feature detection. In recent years, the convolutional neural network (CNN) became the standard method of feature detection since it allowed to optimize thousands of kernels at the same time. However, a limitation of the CNN is that all the kernels are small (usually between 3x3 and 7x7), which limits the receptive field. Another limitation is that feature merging is done via weighted additions and pooling, which cannot be used to merge spatial-domain features with gradient-domain features since they are not located at the same pixel coordinate. The objective of this thesis is to develop electromagnetic (EM) convolutions and Green’s functions (GF) convolutions to be used in Computer Vision and convolutional neural networks (CNN). These new kernels do not have the limitations of the standard CNN kernels since they allow an unlimited receptive field and interaction between any pixel in the image by using kernels bigger than the image. They allow merging spatial domain features with gradient domain features by integrating any vector field. Additionally, they can transform any vector field of features into its least-error conservative field, meaning that the field of features becomes smooth, irrotational and conservative (line-integrable). At first, we developed different symmetrical and asymmetrical convolutional kernel based on EM and GF that are both resolution and rotation invariant. Then we developed the first method of determining the probability of being inside partial edges, which allow extrapolating thin edge features into the full 2D space. Furthermore, the current thesis proves that GF kernels are the least-error gradient and Laplacian solvers, and they are empirically demonstrated to be faster than the fastest competing method and easier to implement. Consequently, using the fast gradient solver, we developed the first method that directly combines edges with saliency maps in the gradient domain, then solves the gradient to go back to the saliency domain. The improvement of the saliency maps over the F-measure is on average 6.6 times better than the nearest competing algorithm on a selected dataset. Then, to improve the saliency maps further, we developed the DSS-GIS model which combines edges with salient regions deep inside the network.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.

Full frame distilled prediction

Teacher imitation

Not calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.

metaresearch head score (Codex)0.000
metaresearch head score (Gemma)0.000
Version: codex-gemma-dda1882f352aValidation status: machine_predicted_unvalidated
Candidate categoriesnone
Consensus categoriesnone
DomainCandidate signal: none · Consensus signal: none
Study designCandidate signal: Simulation or modeling · Consensus signal: Simulation or modeling
GenreCandidate signal: Methods · Consensus signal: none
Teacher disagreement score0.344
Threshold uncertainty score0.866

Codex and Gemma teacher scores by category

CategoryCodexGemma
Metaresearch0.0000.000
Meta-epidemiology (narrow)0.0000.000
Meta-epidemiology (broad)0.0000.000
Bibliometrics0.0000.000
Science and technology studies0.0000.000
Scholarly communication0.0000.000
Open science0.0000.000
Research integrity0.0000.000
Insufficient payload (model declined to judge)0.0000.000

Machine scores (provisional)

The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.

Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.

Opus teacher head0.007
GPT teacher head0.212
Teacher spread0.205 · how far apart the two teachers sit on this one work
Validation statusscore_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it