V-CNN: A Versatile Light CNN Structure For Image Recognition On Resources Constrained Platforms

This paper introduces V-CNN, a versatile structure inspired from previously introduced light models such as NL-CNN. What makes the V-CNN structure efficient is the process of optimizing its hyper-parameters. Consequently, several aspects in proper design for building of efficient (validation accuracy near state of the art while keeping the complexity of the structure) V-CNN structures, are considered and detailed in this paper. While V-CNN includes as particular cases several previously defined models such as NL-CNN and XNL-CNN, it can be better tailored for efficient deployment given a specific dataset. A V-CNN model with 1.5 million parameters obtained 91.55% validation accuracy on CIFAR-10 dataset, surpassing the previous result of 90.60% using the NL-CNN. The V-CNN model offers the following advantages: i) using optimization hints described herein it allows maximal efficiency (good accuracy at low complexity) for a wide variety of datasets; ii) has a low number of layer primitives, thus making easier their specific design for deployment on various TinyML or EdgeAI platforms, including FPGAS.