Optimal Ensemble Classifiers Based Classification for Automatic Vehicle Type Recognition

CNN Classifier’s Robustness Enhancement when Preserving Privacy

“…Our dataset is constructed from a set of 400K vehicle images collected from the pay tolls owned by the Vinci Autoroutes company. We refer readers to the supplementary material or [2,3] for more examples and details. Note that, for this research we intentionally make our dataset challenging by selecting particular samples for which the original classifier has lower confidence.…”

“…Fig. 2 illustrates the architecture, which duplicates the VGG based model of [3] to construct Siamese model [7]. It facilitates implementing the third (features' distance minimization) objective using the mean squared error loss [6].…”

“…Let f xi denotes the CNN features of input x i . Existing model [3] uses the Softmax function [6] to formulate the classification objective (minimize cross entropy) as follows:…”

“…This reduces the reliability of the large-scale GDPR compliance industrial systems. In this paper, we illustrate this problem by exposing a real world application: Automatic Toll Collection (ATC) [2,3]. ATC classifies millions of vehicles every day from the original image, saves images after anonymizing privacy information and later classifies the saved images if necessary.…”

“…Experiments on a challenging vehicle classification dataset show that the proposed method outperforms the competitive methods to achieve better accuracy, lower decision errors and hence higher robustness on anonymization. Our contributions can be summarized as: (a) we present a highly important and contemporary problem for the CNNs related to privacy preserving classification; (b) we propose a novel method to perform privacy preserving vehicle classification [2,3], and (c) we achieved very high performance compared to the other competitive methods.…”