Abstract

Edge detection plays a very crucial role in many image processing, and computer vision applications. Several edge detection methods have been proposed in the literature, which can be categorized into two groups: non learning based methods and learning based methods. The performance of the methods in the first category is not as good as of the methods in the second category. Use of deep convolutional neural networks (DCNNs) has significantly advanced the performance of image edge detection techniques. Most of the existing DCNN techniques are based on the ResNet architecture or on the VGG-16 architecture. The ResNet based techniques exhibit a good edge detection performance at the expense of an extremely high computational complexity due to the use of very large number of convolutional layers. The VGG-16 based techniques have much lower complexity, however, their performance is inferior to that of ResNet based techniques. This inferior performance is due to the fact that the spatial resolution of the feature maps in the last layer is very small. Restoring this small spatial resolution to the original image size would result in a blurred output and a poor localization of the edges. In addition, the deeper layers of the VGG16-based techniques are not able to learn some of the information captured by the initial layers. In this thesis, deep convolutional neural networks based on the VGG-16 architecture, with a focus on a reduced complexity and a performance that is comparable or superior to those of the other existing edge detection techniques, are developed. The thesis has two parts.

In the first part, the idea of residual learning is introduced for the first time in a VGG-16 architecture for the task of edge detection with a view to improve the performance of the existing VGG-16 based networks and also to reduce the complexity. The idea of residual learning enables the network to progressively increase the spatial resolution of the maps as the features extraction process is moved from the shallow layers to the deeper layers of the network through an appropriate use of transposed convolutions, and the use of smaller number and larger size filters in the deeper layers. The proposed network is experimented on different datasets and is shown to outperform most of the existing techniques. At the same time, the complexity of the proposed network in terms of the number of parameters is lower than that of all the other existing edge detection techniques.

Even though the complexity of the technique proposed in the first part is lower than that of all of the other networks, it is still high and its performance is lower than that of a couple of the other existing techniques, one of which is VGG-16 based and the other is ResNet based. Hence, it is important to develop an edge detection convolutional network having a complexity that is still lower than that of the network proposed in the first part, but without lowering its performance. With this objective in mind, in the second part, we develop deep residual convolutional neural networks based on the VGG-16 architecture. The objective of reduced complexity of the networks is achieved through the use of fire modules which results in increasing the depth of the proposed networks. Also, the use of residual learning allows to maintain or even improve the performance of the networks. The objectives of the proposed networks are validated by conducting experiments employing two different datasets and the proposed networks are shown to outperform all the existing techniques in terms of the edge detection accuracy and complexity.