Deep convolutional neural networks for regular texture recognition

View article
Loading...
PeerJ Computer Science

Main article text

 

Introduction

Methods

Regular texture database

Deep convolutional neural networks

where l represents the lth layer, * is a convolution operation (filter), ωl is the weight matrix, bl is the vector (bias) and σ is the nonlinear activation function.

where i is the iteration index, ε is the learning rate, λ is the weight decay, v is the momentum variable, r is the momentum weight and Lω|ωiDi is the average error over the ith batch D of the derivative of the objective function with respect to ω, evaluated at wi. The cross-entropy loss function was: L(ω)=jcyjclogfc(gj), where yjc denotes the label for an image with indexed j and class c; fc (gj) is the prediction probability of class c for image g.

Regular textures classification

Experimental setup

where εi is the learning rate for the current iteration i; ε0 is the initial learning rate specified as an argument to SGD and λ is the decay rate which is greater than zero. We investigated the effect of a set of different values of the learning rate decay for CNNs in the Results section. We also compared the effect of using data augmentation features such as the shift, flip and rotation from Keras.

Results

Discussions

Conclusions

Supplemental Information

Irregular textures - Part 1.

DOI: 10.7717/peerj-cs.869/supp-1

Irregular textures - Part 2.

DOI: 10.7717/peerj-cs.869/supp-2

Regular textures - Part 1.

DOI: 10.7717/peerj-cs.869/supp-3

Regular textures - Part 2.

DOI: 10.7717/peerj-cs.869/supp-4

Code file.

This code implement the several convolutional neural networks for classifying regular textures from irregular textures.

DOI: 10.7717/peerj-cs.869/supp-5

Code for CNN features.

DOI: 10.7717/peerj-cs.869/supp-6

SVM classifier code.

DOI: 10.7717/peerj-cs.869/supp-7

Additional Information and Declarations

Competing Interests

Xizhi Li is employed by Henan Highway Development Co. LTD. The authors declare that they have no competing interests.

Author Contributions

Ni Liu conceived and designed the experiments, performed the experiments, analyzed the data, performed the computation work, prepared figures and/or tables, authored or reviewed drafts of the paper, and approved the final draft.

Mitchell Rogers performed the experiments, performed the computation work, prepared figures and/or tables, authored or reviewed drafts of the paper, and approved the final draft.

Hua Cui analyzed the data, authored or reviewed drafts of the paper, and approved the final draft.

Weiyu Liu performed the experiments, authored or reviewed drafts of the paper, and approved the final draft.

Xizhi Li performed the computation work, authored or reviewed drafts of the paper, and approved the final draft.

Patrice Delmas conceived and designed the experiments, authored or reviewed drafts of the paper, and approved the final draft.

Data Availability

The following information was supplied regarding data availability:

The code files for implementing convolutional neural networks. Fisher vector pooling and SVM solver and the regular textures dataset are available in the Supplemental Files and at GitHub: https://github.com/NiLiu64/Regular-texture-recognition.

Funding

This work was supported by the National Natural Science Foundation of China (Nos. 61806023 and 61572083), the Henan Provincial Department of Transportation Science and Technology Project (No. 2021G8). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

6 Citations 2,183 Views 466 Downloads