“…The filters used in this approach are built by maximising the statistical independence of the filter responses on a set of sub‐windows extracted from natural images by independent component analysis. To increase the performance of this descriptor, an ensemble, as in [33], is built by varying the parameters of the approach: the filter size (size ∈ {3, 5, 7, 9, 11}); and the threshold used for binarising (th ∈ {−9, −6, −3, 0, 3, 6, 9}). In total, the ensemble is built with 35 SVMs combined by sum rule, with each SVM trained using a different feature vector extracted with a possible ( size , th ) combination of the parameters of BSIF.…”