Multipath feature recalibration DenseNet for image classification
- PDF / 1,141,036 Bytes
- 10 Pages / 595.276 x 790.866 pts Page_size
- 71 Downloads / 280 Views
ORIGINAL ARTICLE
Multipath feature recalibration DenseNet for image classification Bolin Chen1 · Tiesong Zhao1 · Jiahui Liu1 · Liqun Lin1 Received: 5 March 2020 / Accepted: 28 August 2020 © Springer-Verlag GmbH Germany, part of Springer Nature 2020
Abstract Recently, deep neural networks have demonstrated their efficiency in image classification tasks, which are commonly achieved by an extended depth and width of network architecture. However, poor convergence, over-fitting and gradient disappearance might be generated with such comprehensive architectures. Therefore, DenseNet is developed to address these problems. Although DenseNet adopts bottleneck technique in DenseBlocks to avoid relearning feature-maps and decrease parameters, this operation may lead to the skip and loss of important features. Besides, it still takes oversized computational power when the depth and width of the network architecture are increased for better classification. In this paper, we propose a variate of DenseNet, named Multipath Feature Recalibration DenseNet (MFR-DenseNet), to stack convolution layers instead of adopting bottleneck for improving feature extraction. Meanwhile, we build multipath DenseBlocks with Squeeze-Excitation (SE) module to represent the interdependencies of useful feature-maps among different DenseBlocks. Experiments in CIFAR-10, CIFAR-100, MNIST and SVHN reveal the efficiency of our network, with further reduced redundancy whilst maintaining the high accuracy of DenseNet. Keywords DenseNet · Image classification · Multipath DenseBlocks · Feature recalibration · MFR-DenseNet
1 Introduction In recent years, convolutional neural networks (CNNs) are ongoing towards an extended depth and width of the network architecture due to the improvement of hardware devices. From AlexNet [1] to VGGNet [2] as well as GoogleNet [3], Highway [4] and ResNet [5], both the depth and the width of CNNs have continued to increase, which can facilitate to achieve excellent performance in image classification tasks. This research is supported by the National Natural Science Foundation of China (grant 61671152, 61901119). * Liqun Lin [email protected] Bolin Chen [email protected] Tiesong Zhao [email protected] Jiahui Liu [email protected] 1
Fujian Key Lab for Intelligent Processing and Wireless Transmission of Media Information, College of Physics and Information Engineering, Fuzhou University, Fuzhou 350116, Fujian, China
However, related works [3, 6, 7] demonstrate that such drawbacks as poor convergence, over-fitting and gradient disappearance might be generated with comprehensive network structures. To alleviate these problems and encourage feature propagation/reuse, DenseNet [8] allows layers access to feature-maps from all of its preceding layers. DenseNet [8], one of the existing deepest CNNs, is mainly composed of DenseBlocks (extracting and propagating feature-maps) and transition layers (connecting two adjacent DenseBlocks). Although DenseNet has achieved remarkable results due to dense connectivity and deep stru
Data Loading...
