A low-complexity residual deep neural network for image edge detection

In many computer vision and image processing applications, edge detection is a crucial step. However, in real time applications both the accuracy and the time complexity of edge detection are very important. Even though the performance of image edge detection techniques can be greatly improved through the use of deep convolutional neural networks (DCNNs), such networks generally result in a significantly increased computational complexity. Several techniques have been developed for edge detection based on the VGG16 network, since the convolutional layers of the networks of such schemes have fewer parameters than those of the existing residual networks such as ResNet50. However, their performance is inferior to that of the residual techniques and their computational complexity is still very large. In this article, a DCNN based on the VGG-16 architecture, with a focus on a significantly reduced complexity but with a performance that is comparable or superior to those of all the other existing edge detection techniques, is proposed. The objective of significantly reduced complexity of the network is achieved through the use of fire modules so much so that it is possible to increase the depth of the network while keeping its character of low complexity. This along with the use of residual learning allows to maintain or even to improve the performance of the network. The objectives of the proposed scheme are validated by conducting experiments employing two different datasets.