Neuroevolution Architecture Backbone for X-ray Object Detection

Most of the developments in Convolutional Neural Networks come from creating innovative designs in the architecture and hyperparameters. Architecture designs often grow in depth and usually outperform their predecessors. However, these deep networks are not optimized for a specific application, especially with limited datasets such as X-ray dataset. Moreover, training these networks require considerable GPU resources. YOLOv3 is an object detection network iterated to include the recent developments in the deep learning research field. Thus, it also has a deep backbone network similar to Residual Network. The developments incorporated are effective in general datasets and benchmarks but needlessly big for a specialized application. Hence, this paper proposes methods to create a suitable backbone for YOLOv3 with Neuroevolution. Neuroevolution is an optimization algorithm based on genetic algorithm, which does not require gradients to optimize. Using Neuroevolution, a small network for limited dataset can be designed without human expertise. Experimental results show that it can compete with the larger networks in terms of accuracy.