Reduced dimensional features for object recognition
Reduced dimensional features for object recognition
Dosyalar
Tarih
2018-07-27
Yazarlar
Keser, Reyhan Kevser
Süreli Yayın başlığı
Süreli Yayın ISSN
Cilt Başlığı
Yayınevi
Institute of Informatics
Özet
Object recognition is one of the substantial problems of computer vision area. Traditional solutions consist of feature based object recognition techniques. Hence, there are many studies which are proposed feature detection and description methods. Object recognition can be performed with high accuracy thanks to these robust features. However, these features suffer from their high dimensional structure, in other words "curse of dimensionality". Hence, dimensionality reduction of the feature vectors is quite studied and methods that reduce computational load are proposed, in the literature. In this thesis, dimensionality reduction of visual features using autoencoders is proposed. And, the effect of dimensionality reduction of visual features are investigated on object recognition task. For this purpose, three well-known feature vectors are selected which are Histogram of Oriented Gradients (HOG), Scale Invariant Feature Transform (SIFT) and Speeded-Up Robust Features (SURF). To conduct experiments, three subsets of Caltech-256 dataset images are designed and HOG, SIFT and SURF feature vectors are obtained from these subsets. Dimensionality of these feature vectors are reduced to half using autoencoders. Then, object recognition is tested with original and reduced dimensional vectors with three different distance measures. Autoencoders which are unsupervised neural network algorithms, are selected for dimensionality reduction of feature vectors since autoencoders can capture nonlinear relationship in data, provide trained model for new inputs and do not need labels. Also, Principal Component Analysis (PCA) is used for dimensionality reduction of these feature vectors for comparison, since PCA is commonly used for dimensionality reduction of these vectors in the literature. Moreover, experiments using the proposed method and PCA, are repeated on images with noise and results are reported. The results show that object recognition accuracies are improved owing to dimensionality reduction. This shows that unnecessary features and noise are eliminated by dimensionality reduction. In addition to this, dimensionality reduction provides memory and time efficiency.
Açıklama
Thesis (M.Sc.) -- İstanbul Technical University, Institute of Informatics, 2018
Anahtar kelimeler
big data,
büyük veri,
object recognition,
nesne tanıma,
vector analysis,
vektör analizi