On the spectral-based reduction of the computational demand for the prediction phase in image classification with artificial neural networks

Computer vision abilities of autonomous systems have a significant computational demand on the underlying hardware. We research on methods to reduce the amount of data provided to Artificial Neural Networks (ANN)-based image classification. We consider transformation techniques, as many visual sensors inherently provide hardware components for the Discrete Cosine Transform (DCT). The focus in this paper is on a fast prediction phase, at the cost of memory and a more sophisticated training phase. In particular, we partition data in the frequency domain, define the problem by mathematical programming and train a set of distinct ANNs. An online algorithm selects the smallest feasible ANN in the prediction phase.