Object Detection Using Deep Convolutional Generative Adversarial Networks Embedded Single Shot Detector with Hyper-parameter Optimization

Abstract

Itis a challenging task to identify optimal network configurations for large-scale deep neural networks with cascaded structures. In this research, we propose a hybrid end-to-end model by integrating Deep Convolutional Generative Adversarial Network (DCGAN) with Single Shot Detector (SSD), for undertaking object detection. We subsequently employ the Particle Swarm Optimization (PSO) algorithm to conduct hyperparameter identification for the DCGAN-SSD model. The detected class labels as well as salient regional features are then used as inputs for a Long Short-Term Memory (LSTM) network for image description generation. Evaluated with a video data set in the wild, the empirical results indicate the efficiency of the proposed PSO-enhanced DCGAN-SSD object detector with respect to object detection and image description generation.