Optimization Of YOLOv3 Object Detection With Dilated Convolutional Layer For Non-GPU Devices

Deep Convolutional Neural Networks(CNNs)is having a fundamental role in the evolution of deep learning.The utilization of CNNs to image recognition projects has paved the way for fruitful results.Having achieved this momentum,deep learning was drawn-out to other computer vision functions like object detection,object tracking etc.CNNs preserve the input-spatial information of the input signals which is very pragmatic for functions like image recognition.Nevertheless,a CNN is in need of tremendous complex computations and very importantly,storage.This prevented the implementation of deep learning layouts to resource-limited devices such as laptops and mobile phones.To circumvent such limitations,a number of approaches were tested to accelerate deep models and thus reducing the memory requirement.Furthermore,a lot of the approaches were applicable either to smaller networks or only for image recognition tasks.The spotlight of this thesis is to cut down on the number of computations in CNN based object detection models during inference.Diverse approaches to solve the problem were evaluated and dilated convolution with the effective receptive field was found to be effective.Dilation is a method of removing redundant filters or neurons and thereby reduce the network complexity.We use the term "dilated convolution",to clarify that no "dilated filter" is constructed or represented.The dilated convolution operator is tailored to utilize parameters filtered in a number of different ways.This application can be reinforced by the input of different filter ranges,with the help of dilation factors.The qualitative and quantitative results of this thesis show the performance of the proposed method for varying degrees of dilation on Tiny YOLOv3.