VRES-CNN: A Tiny Convolutional Image Classifier with Versatile Choice of Hyperparameters
Radu Dogaru, Ioana Dogaru · 2024
A novel lightweight convolutional neural network architecture is proposed, denotes as VRES-CNN. Unlike most of the actual lightweight architectures VRES-CNN provides a versatile set of hyper-parameters associated with variable number of macro-blocks. Moreover, the macro-block definition is rather simple and allows very good accuracies by emulating nonlinear convolutions. The use of residual connections, dropout, batch normalization and separable convolutions ensure very good accuracies with a relatively small complexity of the model. With a proper choice of the hyper-parameters, it is demonstrated that models with less than 100 kilo parameters and near state-of-the-art accuracy can be identified for a wide range of datasets (with low or high input image resolutions). In terms of accuracy performance, it is shown that VRES-CNN performs in the state-of-the-art range and much better than a recently proposed lightweight architecture (EtinyNet) for the same or smaller model size.