UPM Institutional Repository

Plant identification using combination of fuzzy c-means spatial pyramid matching, gist, multi-texton histogram and multiview dictionary learning


Safa, Soodabeh (2016) Plant identification using combination of fuzzy c-means spatial pyramid matching, gist, multi-texton histogram and multiview dictionary learning. Doctoral thesis, Universiti Putra Malaysia.


Plants identification has become a significant and incentive research area. It is estimated that about half of the world's plant species are still not identified. Making a detailed knowledge of the identity and geographical distribution of plants is required for an effective agricultural biodiversity. Most of the existing plant identification methods are based on both the global shape features and the intact plant leaves. However, for the non-intact leaves such as the deformed, partial and overlapped leaves that largely exist in practice, the global shape features are not efficient and these methods are not applicable.The dried leave parts and noise can degrade identification results and affect the quality of the extracted features which lead to poor classification results. Furthermore, feature extraction methods based on global features such as shape, color and texture do not lead to accurate identification since they cannot adapt to changing environment. In the real world, leaf images can be simply affected by light, position, and size. To overcome this problem, in recent years, researchers obtained some achievements with combination of invariant local features such as Scale Invariant Feature Transform (SIFT) with global feature of leaf images. Beside that, classic bag of visual words algorithm (BoVW) is based on kmeans clustering and every SIFT feature belongs to one cluster and it leads to decreasing classification results. Moreover with simple concatenating features, classification results are not optimal. It is crucial to integrate these heterogeneous features to create more accurate and robust classification results than using each individual type of features. This study first starts with some preprocessing phases for images with dried and damaged parts in leaves, that applies on images while finding leaf as region of interest (ROI) with Otsu's method. For next, instead of k-means clustring, Fuzzy cmeans clustering is combined with Spatial Pyramid Matching image representation to improve the accuracy of classification results. The Fuzzy c-means clustering improved the accuracy of classification task to 40.53%. In the next phase, the local SIFT descriptor is augmented with two global descriptors. One descriptor contains texture and color called Multi-Textron Histogram (MTH) and improved classification results by second level of discrimination for leaves with similar color and shape. Second one is gist from global features of leaf images. gist descriptor is based on spatial layout of colors, orientation and principal texture. The combination of gist, MTH and SIFT features increased the performance of image identification and showed 49% accuracy. Moreover, instead of concatenating feature vectors together and send to classifier, sparse coding and dictionary learning methods are used and instead of considering all features as one view (visual feature), K-SVD algorithm that is one of the famous algorithms for sparse representation is optimized and developed to multi-view model.The experimental results prove that the proposed methods has improved accuracy by 53.77% compared to concatenating features and classic K-SVD dictionary learning model as well.

Download File

FSKTM 2016 11 IR.pdf

Download (1MB) | Preview

Additional Metadata

Item Type: Thesis (Doctoral)
Subject: Plants - Identification
Subject: Fuzzy logic
Subject: Fuzzy systems
Call Number: FSKTM 2016 11
Chairman Supervisor: Associate Professor Fatimah Khalid, PhD
Divisions: Faculty of Computer Science and Information Technology
Depositing User: Haridan Mohd Jais
Date Deposited: 15 Nov 2018 01:13
Last Modified: 15 Nov 2018 04:53
URI: http://psasir.upm.edu.my/id/eprint/65923
Statistic Details: View Download Statistic

Actions (login required)

View Item View Item