Lightweight residual densely connected convolutional neural network
- PDF / 1,198,521 Bytes
- 18 Pages / 439.642 x 666.49 pts Page_size
- 3 Downloads / 214 Views
Lightweight residual densely connected convolutional neural network Fahimeh Fooladgar1 · Shohreh Kasaei1 Received: 26 September 2019 / Revised: 8 June 2020 / Accepted: 15 June 2020 / © Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract Extremely efficient convolutional neural network architectures are one of the most important requirements for limited-resource devices (such as embedded and mobile devices). The computing power and memory size are two important constraints of these devices. Recently, some architectures have been proposed to overcome these limitations by considering specific hardware-software equipment. In this paper, the lightweight residual densely connected blocks are proposed to guaranty the deep supervision, efficient gradient flow, and feature reuse abilities of convolutional neural network. The proposed method decreases the cost of training and inference processes without using any special hardware-software equipment by just reducing the number of parameters and computational operations while achieving a feasible accuracy. Extensive experimental results demonstrate that the proposed architecture is more efficient than the AlexNet and VGGNet in terms of model size, required parameters, and even accuracy. The proposed model has been evaluated on the ImageNet, MNIST, Fashion MNIST, SVHN, CIFAR-10, and CIFAR-100. It achieves state-of-the-art results on Fashion MNIST dataset and reasonable results on the others. The obtained results show the superiority of the proposed method to efficient models such as the SqueezNet. It is also comparable with state-of-the-art efficient models such as CondenseNet and ShuffleNet. Keywords Image classification · Convolutional neural networks · Deep learning · Efficient architecture
1 Introduction In the last decades, Convolutional Neural Networks (CNNs) have changed the landscape of visual recognition tasks such as image classification [17, 18, 20] and semantic segmentation [13, 33, 34]. These models need large training datasets with high-end GPU devices to learn Shohreh Kasaei
[email protected] Fahimeh Fooladgar [email protected] 1
Department of Computer Engineering, Sharif University of Technology, Tehran, Iran
Multimedia Tools and Applications
a large number of parameters via a high number of computational operations. However, the most important issues in CNN models are the hardware and the burden of computational cost. Yet, complex CNNs [20] with high resource demands have been proposed to increase the accuracy. Besides, the ultra-deep CNN models have further increased the depth of networks from 8 layers (AlexNet) [28] to more than one thousand layers (ResNet) [17]. The general idea of CNNs has proceeded through deeper and more complex networks to boost the performance in terms of the model’s accuracy. But, the efficiency of networks in terms of model’s size, inference speedup, and computational costs have been rarely inspected. Deploying CNN models on applications with embedded platforms (such as autonomous driving
Data Loading...