Primjena dubokog učenja u prepoznavanju umjetničkog stila iz digitalnih slika

master's thesis

• Title (english): Deep Learning Approaches to Art Style Recognition in Digital Images
• Author: Andrea Gradečak
• Mentor: Manfred Jaeger (mentor); ISVU oznaka: MJ203
• Committee member: Manfred Jaeger (član povjerenstva); ISVU oznaka: MJ203
• Committee member: Christian Storm Pedersen (član povjerenstva); ISVU oznaka: CS004
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2017-06-20, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Konvolucijske neuronske mreže su dosada najuspješniji model za prepoznavanje slika, ali se do sada nisu značajnije koristile za prepoznavanje stilova u umjetničkim djelima. U ovom diplomskom radu razvili smo konvolucijsku neuronsku mrežu za spomenuti problem eksperimentirajući s raznim segmentima ovog modela. Model ima 7 blokova od kojih svaki sadrži 1 konvolucijski sloj, rectified linear unit i max-pooling sloj. Model ima 32 mape značajki u prvom konvolucijskom sloju, a broj im se udvostručava nakon svakog max-pooling sloja. Na kraju neuronske mreže se nalazi 1 potpuno povezan sloj s 8 neurona te soft-max izlazni sloj. Kao osnovu za naš model i usporedbu rezultata koristili smo
VGG16 mrežu u kombinaciji s metodom Support Vector Machine (SVM) kao klasifikatorom. Kako bismo nadoknadili količinski malen skup podataka i pritom ga uvećali, upotrijebili smo tehniku obrezivanja slike pomoću klizećeg prozora uzimajući maksimalno 10 dijelova originalne slike. Model smo testirali na 3 para umjetničkih stilova i postigli najvišu točnost od 95.3% na skupu Color Field Painting i Magic Realism. Unatoč tomu, rezultat nije nadmašio točnost VGG16 mreže od 97.8%. Na kraju, zaključujemo da učenje konvolucijske neuronske mreže na umjetničkim slikama bez agresivne augmentacije nije uspješnije od korištenja unaprijed naučene mreže i SVM klasifikatora.

Abstract (english)

Convolutional Neural Networks (CNNs) have become state-of-the-art image recognition models, but have not been used to significant effect for art style recognition in fine art paintings. Through incremental experimentation with a number of aspects of CNNs, we have build a model for this task. The model has 7 blocks, containing one convolutional layer with rectified linear units and a max-pooling layer. It has 32 feature maps in the first convolutional layer, which is doubled after each max-pooling layer. At the end of the network there is one fully connected layer with 8 neurons and a softmax output layer. As baseline we have used the VGG16 network with a Support Vector Machine(SVM) as classifier. To compensate for the relatively small size of the dataset, we employed a sliding window cropping technique, taking a maximum of 10 crops from the original image to inflate the dataset. Testing on three pairs of styles we reached the highest test accuracy of 95.3% on Color Field Painting and Magic Realism. However, this was not enough to beat the baseline, that reached a test accuracy of 97.8%. To conclude, we find that without aggressive augmentation, training purely on fine art paintings for style recognition, is not viably better than using a pretrained CNN and an SVM classifier.

• Keywords: prepoznavanje umjetničkog stila; konvolucijska neuronska mreža; prepoznavanje slike; duboko učenje; računalni vid
• Keywords (english): art style recognition; convolutional neural network; image recognition; deep learning; computer vision
• Language: english
• URN:NBN: urn:nbn:hr:168:338904
• Study programme: Title: Computing; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2019-03-19 18:48:21