PNGNET: A CAMOUFLAGE DETECTION NETWORK BASED ON PYRAMID POOLING MODULE

In recent years, recent researchers like Fan et al. have developed methods based on deep learning such as PraNet and SINet, which contributed a lot to the field of camouflage object detection. However, there is a contradiction between the receptive field of feature extraction and the resolution of t...

Full description

Saved in:
Bibliographic Details
Published inScientific Bulletin. Series C, Electrical Engineering and Computer Science no. 1; p. 115
Main Author Guo, Xincheng
Format Journal Article
LanguageEnglish
Published Bucharest University Polytechnica of Bucharest 01.01.2022
Subjects
Artificial neural networks
Camouflage
Feature extraction
Feature maps
Image segmentation
Machine learning
Modules
Object recognition
Online AccessGet full text

Cover

More Information
Summary:In recent years, recent researchers like Fan et al. have developed methods based on deep learning such as PraNet and SINet, which contributed a lot to the field of camouflage object detection. However, there is a contradiction between the receptive field of feature extraction and the resolution of the feature map. To deal with the problem, we adopt the method of integrating the features of different receptive fields with the features of sub-regions. And this paper proposes a new deep neural network for camouflaged detection - PNGNet. By using the Pyramid Pooling Module, the semantic task of the image segmentation of the camouflage object is more accurate and the feature representation ability is enhanced. After times of experiments, the PNGNet outperforms most of the models from the perspective of three representative camouflage detection datasets with higher performance and robustness. The proposed network model takes into account both the receptive field of feature extraction and the resolution of the feature map and provides a new idea for camouflage detection. In future work, the results can also be applied in the medical field of segmenting pneumonia and polyps, image search, field rescue, and other fields.
ISSN:2286-3540