Learning invariant features through topographic filter maps

Kavukcuoglu, Koray; Ranzato, Marc ' Aurelio; Fergus, Rob; LeCun, Yann

doi:10.1109/cvpr.2009.5206545

Cited by 218 publications

(148 citation statements)

References 20 publications

Supporting

Mentioning

146

Contrasting

Order By: Relevance

“…Several works have shown that applying sparse coding to local parts or descriptors of the images can capture higher-level features compared to raw image patches [38,39]. In our work, we divide the image into square patches of fixed size and extract low-level features from each of the patches.…”

Section: Features Extraction and Representationmentioning

confidence: 99%

Discrimination of retinal images containing bright lesions using sparse coded features and SVM

Sidibé

Sadek

Mériaudeau

2015

Computers in Biology and Medicine

View full text Add to dashboard Cite

Diabetic Retinopathy (DR) is a chronic progressive disease of the retinal microvasculature which is among the major causes of vision loss in the world. The diagnosis of DR is based on the detection of retinal lesions such as microaneurysms, exudates and drusen in retinal images acquired by a fundus camera. However, bright lesions such as exudates and drusen share similar appearances while being signs of different diseases. Therefore, discriminating between different types of lesions is of interest for improving screening performances. In this paper, we propose to use sparse coding techniques for retinal images classification. In particular, we are interested in discriminating between retinal images containing either exudates or drusen, and normal images free of lesions. Extensive experiments show that dictionary learning techniques can capture strong structures of retinal images and produce discriminant descriptors for classification. In particular, using a linear SVM with the obtained sparse coded features, the proposed method achieves superior performance as compared with the popular Bag-of-Visual-Word approach for image classification. Experiments with a dataset of 828 retinal images collected from various sources show that the proposed approach provides excellent discrimination results for normal, drusen and exudates images. It achieves a sensitivity and a specificity of 96.50% and 97.70% for the normal class; 99.10% and 100% for the drusen class; and 97.40% and 98.20% for the exudates class with a medium size dictionary of 100 atoms.

show abstract

Section: Features Extraction and Representationmentioning

confidence: 99%

Discrimination of retinal images containing bright lesions using sparse coded features and SVM

Sidibé

Sadek

Mériaudeau

2015

Computers in Biology and Medicine

View full text Add to dashboard Cite

show abstract

“…Figure 1 shows an example map of a 2D topography. In this paper, the function will not be learnt as [7,9,10], and we will keep it fixed for simplicity. Some studies on how to estimate the topography function can be found in [8,11,25].…”

Section: Modelmentioning

confidence: 99%

“…In a recent work, Kavukcuoglu et al have proposed an architecture and a learning algorithm that can learn location-invariant feature descriptors [10]. Those feature descriptors include a bank of filters which are learnt by an improved sparse coding model [3] with SRSS pooling.…”

Section: The Invariance To Translationsmentioning

confidence: 99%

See 1 more Smart Citation

Learning Topographic Representations of Nature Images with Pairwise Cumulant

Wang

Huang

Luo

et al. 2011

Neural Process Lett

View full text Add to dashboard Cite

In this paper, we propose a model for natural images to learn topographic representations and complex cell properties. Different from the estimation of traditional models, e.g., pooling the outputs of filters in neighboring regions, our method maximizes a simple form of binary relations between two adjacent complex cells-"pairwise cumulant", which contains the favorable nonlinearity as high order cumulant, and can exploit the "sparseness" and "correlation" of cells in primary visual cortex. By means of choosing nonlinearity properly, our model is related to cumulant-based ICA model, and the derived fixed-point algorithm is close to the well-known FastICA algorithm. The local convergence analysis proves that our fixed-point algorithm is cubic convergence and experiments on nature images show its high efficiency than traditional algorithms. Besides, simulations demonstrate the effectiveness of our model in capturing nonlinear dependencies among these neighboring complex cells. The learnt filters preserve properties of complex cells, and their orientation, spatial frequency and location change smoothly over the topographic map. In addition, these learnt filters can be used as feature descriptors. They produce features that are invariant to object transformations, and achieve better results than traditional models on digit recognition tasks.

show abstract

“…The first is the structured dictionary learning, where the interaction between the dictionary atoms is also learned/imposed. Common structures are tree structures [39] and grid structures [46]. The second variation is multiscale dictionary learning, extending the basic dictionary learning scheme to consider different patch sizes [68].…”

Section: Learning Overcomplete Dictionariesmentioning

confidence: 99%