Proceedings of 2nd International Multi-Disciplinary Conference Theme: Integrated Sciences and Technologies, IMDC-IST 2021, 7-9 September 2021, Sakarya, Turkey

Research Article

Hierarchical Feature Mining for Image Classification and Segmentation

Download279 downloads
  • @INPROCEEDINGS{10.4108/eai.7-9-2021.2315306,
        author={Hayder  Yousif and Zahraa  Al-Milaji},
        title={Hierarchical Feature Mining for Image  Classification and Segmentation},
        proceedings={Proceedings of 2nd International Multi-Disciplinary Conference Theme: Integrated Sciences and Technologies, IMDC-IST 2021, 7-9 September 2021, Sakarya, Turkey},
        publisher={EAI},
        proceedings_a={IMDC-IST},
        year={2022},
        month={1},
        keywords={dcnn supervised learning unsupervised learning k-means graph- cut},
        doi={10.4108/eai.7-9-2021.2315306}
    }
    
  • Hayder Yousif
    Zahraa Al-Milaji
    Year: 2022
    Hierarchical Feature Mining for Image Classification and Segmentation
    IMDC-IST
    EAI
    DOI: 10.4108/eai.7-9-2021.2315306
Hayder Yousif1,*, Zahraa Al-Milaji1
  • 1: Basra Engineering Technical College, Southern Technical University, Iraq
*Contact email: hayder.yaqoob@stu.edu.iq

Abstract

In this paper, supervised and unsupervised learning are used to classify and segment two different types of images. Specifically, first, we use our deep learning classification model to develop a fast and accurate scheme for feature extraction. We design a DCNN model that is 18 times faster than the state-of-the-art AlexNet at a small loss of accuracy. Second, different classifiers have been trained on the features that are extracted from the fully connected layer of the deep convolutional neural networks (DCNN). Finally, we perform K-means-graph-cut segmentation and qual- iteratively compare with supervised segmentation results. Image-level and region-level classification has been studied by utilizing natural scenes and biomedical image processing, respectively. Our experimental results demonstrate that the proposed method outperforms existing methods on a biomedical image dataset.