Emotion recognition in blurred images with local features and machine learning

Følelsesgenkendelse i slørrede billeder ved brug af lokale billedtræk og maskinindlæring

Kristensen, Rasmus Lyngby

4. term

Vision, Graphics and Interactive Systems, Master

2014

2014-06-04

130

Ansigtsudtryksgenkendelse (FER) kan gøre fremtidige menneske–maskine-grænseflader mere naturlige og lydhøre. Dette studie undersøger, hvordan man kan genkende udtryk, når kameraets billeder er uskarpe, fx på grund af bevægelse eller forkert fokus. Det evaluerer en slør-robust lokal trækkarakteristik kaldet Local Frequency Descriptor (LFD) og foreslår at bruge Spectral Clustering (SC) til at reducere dimensionaliteten af lokale træk. LFD sammenlignes med en anden slør-robust deskriptor, Local Phase Quantization (LPQ), samt den udbredte metode Local Binary Patterns (LBP). SC sammenlignes med Principal Component Analysis (PCA) til dimensionalitetsreduktion. Genkendelsesnøjagtighed måles med Support Vector Machines (SVM), en standard klassifikationsmetode i maskinlæring. Resultaterne viser, at LPQ ser ud til at give bedre genkendelsesnøjagtighed end LFD. SC giver generelt lignende eller en smule bedre resultater end PCA, men er mere beregningstung.

Facial Expression Recognition (FER) can make future human–machine interfaces more natural and responsive. This study focuses on recognizing expressions when camera images are blurred, for example by motion or being out of focus. It evaluates a blur-invariant local feature descriptor called Local Frequency Descriptor (LFD) and proposes using Spectral Clustering (SC) to reduce the dimensionality of local features. LFD is compared with another blur-invariant descriptor, Local Phase Quantization (LPQ), and with the widely used Local Binary Patterns (LBP). SC is compared with Principal Component Analysis (PCA) for dimensionality reduction. Recognition accuracy is measured using Support Vector Machines (SVM), a standard machine learning classifier. The results indicate that LPQ appears to provide better recognition accuracy than LFD. SC generally yields similar or slightly better performance than PCA, but at a higher computational cost.

[This abstract was generated with the help of AI]

Blur ; Computer Vision ; Machine Learning ; Local Features ; Local Binary Patterns ; Local Directional Pattern variance ; Local Phase Quantization ; Local Frequency Descriptor ; Spectral Clustering ; Principal Component Analysis ; Segmentation ; Face ; Facial Expression ; Emotion ; Recognition ; Robustness ; Blur robustness

Download
View record in AAU Student Projects