Research On Fast R-CNN For Visual Object Detection

Over the last decade,object detection has emerged as an important area of research for computer vision.In this regard,some novel and impressing CNN architectures have been proposed.For this work Fast R-CNN and VGGNet architectures have been incorporated for the given task of visual object detection.Different proposed methods have been exercised and empirically it is realized that Fast R-CNN architecture exhibits considerable performance for the given task.In particular for given problem,emphasis has been rendered on visual object Capra Falconeri(Markhor),in different challenging environment and background images.The most issue for object detection using deep learning techniques is to save the training and testing time without sacrificing its accuracy.However,existing Fast R-CNN technique provide good accuracy but it suffers from more training and testing time that limits the overall performance of system.In this thesis,a new dataset is created named as Markhor-VI by various data augmentation methods such as translation,rotation,scaling and so on.Then we have modified existing Fast R-CNN technique by transfer learning by various methods from layer 1 to layer 5.It is found that convolutional object detection is still evolving as a technology,despite outranking other object detection methods.It is also found that transfer learning can save a lot of resources if fine tuning is performed in a particular way.