In this study, we attempted to develop a method for accelerating parameter optimization of an object detector ensemble over large image datasets by using simulated annealing. We propose a novel sampling-based evaluation method that considers the minimum portion of the dataset required in each iteration to maintain solution quality. This approach can be considered a noisy evaluation of the energy. The sample sizes required during the search process are theoretically determined by adapting the convergence results for noisy evaluation. To determine applicability, we prepared and optimized two ensembles for diabetic retinopathy pre-screening based on microaneurysm detection with convolutional neural network-based and traditional object detectors. Our experimental results indicate that the proposed sampling-based evaluation method substantially reduced the computational time required for optimizing the parameters of the ensembles while preserving solution quality.

Ensemble-based systems are primarily analyzed on how the accuracy of the ensemble depends on that of its members. In this paper, we extend this model with adding a natural constraint regarding a time limit within which the ensemble should make the decision. For this aim, we consider both the execution time and accuracy of each member. Then, we solve the problem on how to find the most accurate ensemble, where the sum of the execution times of its members remains below the limit. As a decision rule, we analyze a majority voting-based one generalized to be applicable in single object detection scenarios. The optimization task leads to a non-separable Knapsack problem, which is addressed using stochastic considerations. The proposed methodology is also validated experimentally for the localization of the optic disc in retinal images.

Generating ensembles from multiple individual classifiers is a popular approach to raise the accuracy of the decision. As a rule for decision making, majority voting is a usually applied model. In this paper, we generalize classical majority voting by incorporating probability terms pn,k to constrain the basic framework. These terms control whether a correct or false decision is made if k correct votes are present among the total number of n. This generalization is motivated by object detection problems, where the members of the ensemble are image processing algorithms giving their votes as pixels in the image domain. In this scenario, the terms pn,k can be specialized by a geometric constraint. Namely, the votes should fall inside a region matching the size and shape of the object to vote together. We give several theoretical results in this new model for both dependent and independent classifiers, whose individual accuracies may also differ. As a real world example, we present our ensemble-based system developed for the detection of the optic disc in retinal images. For this problem, experimental results are shown to demonstrate the characterization capability of this system. We also investigate how the generalized model can help us to improve an ensemble with extending it by adding a new algorithm.