Contemporary Vision and Pattern Recognition problems such as face recognition, fingerprinting identification, image categorization, and DNA sequencing often have an arbitrarily large number of classes and properties to consider. To deal with such complex problems using just one feature descriptor is a difficult task and feature fusion may become mandatory. Although normal feature fusion is quite effective for some problems, it can yield unexpected classification results when the different features are not properly normalized and preprocessed. Besides it has the drawback of increasing the dimensionality which might require more training data. To cope with these problems, this paper introduces a unified approach that can combine many features and classifiers that requires less training and is more adequate to some problems than a naïve method, where all features are simply concatenated and fed independently to each classification algorithm. Besides that, the presented technique is amenable to continuous learning, both when refining a learned model and also when adding new classes to be discriminated. The introduced fusion approach is validated using a multi-class fruit-and-vegetable categorization task in a semi-controlled environment, such as a distribution center or the supermarket cashier. The results show that the solution is able to reduce the classification error in up to 15 percentage points with respect to the baseline.

Abstract

No abstract

No abstract

Abstract-In this paper, we present an algorithm to detect the presence of diabetic retinopathy (DR) related lesions from fundus images based on a common analytical approach that is capable of identifying both red and bright lesions without requiring specific pre-or post-processing. Our solution constructs a visual word dictionary representing points of interest (PoIs) located within regions marked by specialists that contain lesions associated with DR and classifies the fundus images based on the presence or absence of these PoIs as normal or DRrelated pathology. The novelty of our approach is in locating DR lesions in the optic fundus images using visual words that combines feature information contained within the images in a framework easily extendible to different types of retinal lesions or pathologies and builds a specific projection space for each class of interest (e.g. white lesions such as exudates or normal regions) instead of a common dictionary for all classes. The visual words dictionary was applied to classifying bright and red lesions with classical cross-validation and cross dataset validation to indicate the robustness of this approach. We obtained an AUC of 95.3% for white lesion detection and an AUC of 93.3% for red lesion detection using 5-fold cross-validation and our own data consisting of 687 images of normal retinae, 245 images with bright lesions, 191 with red lesions and 109 with signs of both bright and red lesions. For cross dataset analysis, the visual dictionary also achieves compelling results using our images as the training set and the RetiDB and Messidor images as test sets. In this case, the image classification resulted in an AUC of 88.1% when classifying the RetiDB dataset and in an AUC of 89.3% when classifying the Messidor dataset, both cases for bright lesion detection. The results indicate the potential for training with different acquisition images under different setup conditions with a high accuracy of referral based on the presence of either red or bright lesions or both. The robustness of the visual dictionary against image quality (blurring), resolution, and retinal background, makes it a strong candidate for diabetic retinopathy screening of large, diverse communities with varying cameras and settings and levels of expertise for image capture.