Robust object detection in the wild via cascaded DCGAN

Dinakaran, Ranjith Kumar (2022) Robust object detection in the wild via cascaded DCGAN. Doctoral thesis, Northumbria University.

Preview

Text (Doctoral thesis) dinakaran.ranjith_phd(15029116).pdf - Submitted Version Download (3MB) | Preview

Abstract

This research deals with the challenges of object detection at a distance or low resolution in the wild. The main intention of this research is to exploit and cascade state-of-the-art models and propose a new framework for enabling successful deployment for diverse applications. Specifically, the proposed deep learning framework uses state-of-the-art deep networks, such as Deep Convolutional Generative Adversarial Network (DCGAN) and Single Shot Detector (SSD). It combines the above two deep learning models to generate a new framework, namely DCGAN-SSD. The proposed model can deal with object detection and recognition in the wild with various image resolutions and scaling differences. To deal with multiple object detection tasks, the training of this network model in this research has been conducted using different cross-domain datasets for various applications. The efficiency of the proposed model can further be determined by the validation of diverse applications such as visual surveillance in the wild in intelligent cities, underwater object detection for crewless underwater vehicles, and on-street in-vehicle object detection for driverless vehicle technologies. The results produced by DCGAN-SSD indicate that the proposed method in this research, along with Particle Swarm Optimization (PSO), outperforms every other application concerning object detection and demonstrates its great superiority in improving object detection performance in diverse testing cases. The DCGAN-SSD model is equipped with PSO, which helps select the hyperparameter for the object detector. Most object detectors struggle in this regard, as they require manual effort in selecting the hyperparameters to obtain better object detection. This research encountered the problem of hyperparameter selection through the integration of PSO with SSD. The main reason the research conducted with deep learning models was the traditional machine learning models lag in accuracy and performance. The advantage of this research and it is achieved with the integration of DCGAN-SSD has been accommodated under a single pipeline.

Actions (login required)

View Item

Downloads