VireNet-SSD: object detection model for resource-constrained applications based on self-organized operational neural networks

Abstract Discovering deep learning-based computer vision solutions for use with constrained devices is exceptionally hard, and the trade-offs are often too undermining. Deep learning models are enormous, which makes it challenging to deploy them on constrained platforms. The convolutional neural network is the fundamental framework for majority of the models that are currently in use. However, operational neural networks have recently shown to be a better option to the convolutional equivalents on a variety of tasks due to their heterogeneous nature and greater resemblance to the functioning of biological neurons. The question of whether heterogeneous models could function on constrained devices and be deployed in real time remains a major concern. To address this problem, an object detection model architecture based on a single-shot multi-box detector with self-organized operational neural networks as its backbone was developed, which can perform efficiently on constrained devices such as Raspberry Pi. The resultant backbone architecture was named as VireNet. In contrast to homogeneous conventional deep learning networks that use convolutions, heterogeneous networks were chosen to develop VireNet, which provides a more productive and effective solution. Furthermore, an in-depth explanation of the design space has been provided to aid any future research that is associated with this architectural search. This new approach might mark the very beginning of the use of heterogeneity to address issues on devices with constrained resources.