Diversity of morphological characteristics

Get Complete Project Material File(s) Now! »

Positioning in the literature

Due to the aforementioned needs and challenges, there is now a body of literature in computational vision devoted to fine-grained recognition, including identifying botanical species, and designed for both experts and non-experts. A variety of methods were adapted or introduced to discriminate similar objects and several approaches have made use of human input to improve accuracy. Most of this work is summarized in Chapter 2.
Generally, the baseline scenario was to provide the user with a single estimate; examples include [35, 45, 119, 68, 117, 78, 13, 2]. Other researchers chose to report the k most similar classes [7, 67, 75], where k usually ranges from ten to twenty in order to make it likely the true species is among the k reported ones. Many of the most efficient approaches for fine-grained recognition achieved about 70% accuracy on the first estimate and about 90% on the top-10 estimates while considering challenging and relatively large datasets, i.e., containing more than one hundred species with high
inter-class similarity and intra-class variability; examples include [74, 7, 67, 75]. Of course, retaining the true species while returning a relatively large set of estimates has limited value in real-world applications. Also, focusing only on the few first estimates without achieving near-perfect (human-level) performances could be essentially useless and uninformative for the user. Here, we focus on (1) hierarchical strategies, based on discriminative features, in order to drastically reduce the space of candidate sets of estimates and (2) a new performance criterion: report a subset of species whose expected size is minimized subject to containing the true species with high probability. In particular, we outperform previous methods and achieve near-perfect recognition rates on several leaf images with uniform background; see Chapter 6.

Botanical-based approaches

Domain-specific knowledge is also used to distinguish between similar botanical species.
For example, the vein structure could be very interesting to characterize leaves, but the main challenge is to be able to extract it accurately [73] which generally requires high quality of data. Some previous work combined shape with venation features [93, 87] while other work made use of other morphological leaf information. The authors of [60] investigated local detailed shape of the leaf margin. Typically, they focused on detecting the leaf teeth (on the leaf boundary). Du et al. [36] extracted other properties of the leaf boundary, including aspect ratio, rectangularity, area ratio of convexity, perimeter ratio of convexity, sphericity, circularity and form factor, in order to classify 20 species of plant leaves. Some prior knowledge on simple leaf shape was used to construct a para- metric polygonal leaf template in [22]. Ten models representing classes of leaf shapes were retained and used for classification. Caballero and Aranda [17] used geometric features, including eccentricity and area to reduce the search space while introducing a novel shape-based leaf descriptor. The authors of [4] obtained good results in the ImageCLEF2012 Plant Identification Task 4 by addressing simple and compound leaves separately using many morphological features and a single leaflet analysis for compound leaves. Also, several approaches have exploited specific well-known landmarks and some measurements for leaf retrieval and plant identification. In [64], landmarks were man- ually captured and linear and angular measures were derived from the landmark con- figuration in order to examine relationships between three species of Acer genus. One difficulty in such approaches is the automatic extraction of the landmarks. Recently, Mzoughi et al. [84] introduced an automatic method for detecting different leaf parts and used it in [85] to identify scanned leaf images on white background. Our work is somewhat similar to these morphological approaches in that we also propose to exploit domain knowledge about taxonomy and landmarks in order to build meaningful representation of the object.

READ  The functional requirements of the agricultural sensor board

Segmentation and cluttered background

Work on botanical species identification, including most of those mentioned above, gen- erally deals with leaf images on uniform backgrounds. Often, the Otsu thresholding method [91] was used to extract leaf boundary since the image background is homoge- neous [114, 36, 16, 81, 83, 104]. However, some methods have used more sophisticated algorithms on images with uniform backgrounds such as Expectation-Maximization with post-processing to remove false positive regions [67].
Only few work addressed the problem of identifying leaf images on cluttered back- grounds which is more likely to be the real-world scenario. To tackle this problem, most of them have designed novel segmentation algorithms to overcome the difficulties posed by a natural background. Obviously, isolating green leaves in an overall not less green environment seems like an other more difficult issue. The authors of [112], consid- ered prior shape information and proposed an automatic marker-controlled watershed segmentation method combined with pre-segmentation and morphological operation to segment leaf images with complicated background. Teng et al. [105] proposed to recover the 3D position of a leaf from different cluttered images with close viewpoints.

Table of contents :

1 Introduction
2 Motivations
3 Challenges
4 Positionnement par rapport `a l’´etat-de-l’art
5 Contributions
5.1 Nouvelles repr´esentations de l’objet
5.2 Classification
6 R´esultats
7 Conclusion et perspectives
8 Organisation de la th`ese
1 Introduction 
1 Motivations
2 Challenges
2.1 Lack of data
2.2 Diversity of morphological characteristics
2.3 Image background complexity
3 Positioning in the literature
4 Contributions
4.1 Object representations
4.2 Classification algorithms
4.3 Publications
5 Outline of the thesis
2 Related Work 
1 About fine-grained categorization
1.1 No human intervention
1.2 Human intervention
2 About leaves
2.1 Generic approaches
2.2 Botanical-based approaches
2.3 Segmentation and cluttered background
3 Hierarchical representation and search
4 Class-selective rejection
3 Object representation 
1 Introduction
2 Leaf definition
3 IdKeys
3.1 Motivation
3.2 Hierarchical representation
3.3 Feature extraction
4 Vantage feature frames
4.1 Motivation
4.2 Definition
4.3 Learning the frames
4.4 Detecting the frames
4.5 Learning the features
4.6 Case of leaves
4.7 Case of orchid flowers
5 Summary
4 Classification 
1 Introduction
2 Hierarchy construction
3 Discriminant functions
4 Coarse-to-fine search and likelihood framework
4.1 Coarse-to-fine search
4.2 Likelihood ratios
5 Confidence sets
5.1 Statistical model
5.2 Bayesian network
5.3 Constructing the confidence set
5.4 Relationship to non-Bayesian confidence sets
6 Summary
5 Identification scenarios 
1 Introduction
2 Baseline scenario
3 Final disambiguation
4 Initialization
5 Multiple images
6 Summary
6 Experiments 
1 Datasets
1.1 Swedish leaves
1.2 Flavia leaves
1.3 Smithsonian leaves
1.4 ImageClef leaves
1.5 Orchid flowers
2 Experiments and analyses
2.1 IdKey estimation
2.2 Vantage point detection
2.3 Coarse-To-Fine (CTF) classification
2.4 Confidence sets
3 Summary
7 Conclusion 
1 Summary of contributions
2 Future work
Bibliography

GET THE COMPLETE PROJECT

Related Posts