• 沒有找到結果。

Combining Recognition

CHAPTER 3 PROPOSED system

3.3 Combining Recognition Method

3.3.1 Combining Recognition

In the combining recognition phase, the flower and leaf images of a particular

plant were obtained. Then, the features of the flower and leaf are extracted by the

steps described in sections 3.1 and 3.2. After that, the distances between the query

image and all images in the database are calculated and then the distances are ranked.

The steps are listed below:

1. Get the Top-40 nearest neighbors for the flower and give scores to each rank such

as 1st = 40, 2nd = 39 …, respectively. Next, sum up the scores of the same species.

2. Get the Top-40 nearest neighbors for the leaf and give scores to each rank such as Fig. 21 The flow chart of the proposed method for combining recognition.

Preprocessing Feature Extraction

Combining

Recognition Result Preprocessing Feature

Extraction

1st = 40, 2nd = 39 …, respectively. Next, sum up the scores of the same species.

3. Preserve the species appearing both in step1 and step2. Sum up the scores of the

same species in step1 and step2 as the similarity measure.

4. Rank the similarity measure and return the possible species.

Finally, we can get the most similar species of the query images and the results are

better than using only flower or leaf image.

CHAPTER 4

EXPERIMENTAL RESULTS

In this chapter, experiments are conducted to evaluate the performance of the

proposed method. Firstly, the recognition results of flowers are presented based on

two databases. One is our database of 684 flower images consisting of 24 species and

the other is Zou-Nagy’s [7] database of 612 flower images with 102 species. The

performance of our system in flower recognition will be compared with Zou-Nagy’s

method. Secondly, the recognition results of leaves are presented based on two

databases. One is our database of 960 leaf images consisting of 48 species. The other

is Lee-Chen’s [11] database of 600 leaf images with 60 species. We will compare the

performance between our method and Lee-Chen’s method. Thirdly, the recognition

results of combining system are presented based on two databases. One is our

database consisting of 16 species, including 320 flower images and 320 leaf images.

The other is Saitoh-Kaneko’s [1] database containing 16 species with 320 flower

images and 320 leaf images. The performance of our system will be compared with

Saitoh-Kaneko’s method. Finally, we will compare the performance between our three

recognition methods.

Each flower image in our database is re-scaled to 320x240 pixels. We pick out

one flower image in our database as our query image and consider the remaining 683

flower images as the training data. The results are shown in Table 2. The first row of

Table 2 is measured by returning the Top-5 most similar images to the query image of

the flower recognition method. The second row of Table 2 is measured by returning

the Top-5 most similar species to the query image of the flower recognition method.

Table 2 Performance on our flower database.

Recognition rate (%)

Top-1 Top-2 Top-3 Top-4 Top-5

Number of images

Number of species Similar

images 81.43 89.33 92.25 94.44 96.35 684 24

Similar

species 76.9 93.12 98.25 99.12 99.71 684 24

We also conduct the proposed method on Zou-Nagy’s [7] database collected

from [17]. All images in Zou-Nagy’s database have the same size of 320x240. There

are six images for each species. Some pictures are quite out of focus and the objects

are too small and overlapping. The results are shown in Table 3. We can see that the

processing time of our method is faster than Zou-Nagy’s and the Top-3 recognition

rate (87.6%) is much higher than Zou-Nagy’s (79%) with 8.5 seconds by user’s

rose-curve adjustments before labeling the flower to the class. Although Zou-Nagy’s

method with human help can achieve 93%, it cost much time (10 seconds per flower).

Table 3 Performance comparison between our method and Zou-Nagy’s method using

Zou-Nagy’s database.

Recognition rate (%) Time (s)

Top-1 Top-2 Top-3 Top-4 Top-5 Our method

(similar images) 4.3 76.1 83.8 87.6 90.5 91.3

Our method

(similar species) 4.3 67.3 84.2 92.2 93 93.5

Zou-Nagy’s method

(before labeling) 8.5 52 - 79 - -

Zou-Nagy’s method

(labeling) 10.7 93 - - - -

In our leaf recognition method, each species of leaf includes 40 images; 20 of

them are selected as database images and the remaining are used for testing data. The

results are shown in Table 4 and Table 5. The results of Table 4 are measured by

returning the Top-5 most similar images, and the results of Table 5 are measured by

returning the Top-5 most similar species. The first row of Table 4 and Table 5 is the

performance of the leaves having blooming flowers. The second row of Table 4 and

Table 5 is the performance of the other leaves. The third row of Table 4 and Table 5 is

the performance of all leaves in our database.

Table 4 Performance on our leaf database by returning Top-5 images.

Recognition rate (%)

No. Top-1 Top-2 Top-3 Top-4 Top-5

Number of images

Number of species

1 74.8 86.7 91.5 94 95.8 480 24

2 62.1 76 83.3 87.3 91.3 480 24

3 58.1 71.5 79.9 84.7 87.9 960 48

Table 5 Performance on our leaf database by returning Top-5 species.

Recognition rate (%)

No. Top-1 Top-2 Top-3 Top-4 Top-5

Number of images

Number of species

1. 68.5 90.4 96 98.3 99.4 480 24

2 60 79 88.3 94.6 96.9 480 24

3 52.6 73.1 82.1 89.1 93.2 960 48

We also conduct the proposed method on Lee-Chen’s [4] database. Each image

size of Lee-Chen’s database is 640x480 pixels. Each species in their database includes

15 images; 10 of them are database images and the others are used for testing data.

The results are shown in Table 6. We can see that the recall rate of our method (51.4%)

is much higher than Lee-Chen’s (48.2%) from Table 6. However, Lee and Chen tuned

the weights of features to achieve the optimal recognition rate. Hence, our recognition

rate (70%) is lower than Lee-Chen’s (82.33%) by returning the most similar image.

Nevertheless, our recognition rate can achieve 94.33% by returning the Top-5 most

similar species and it is higher than Lee-Chen’s.

Table 6 Performance comparison between our method and Lee-Chen’s method using

Lee-Chen’s database.

Recognition rate (%)

Top-1 Top-2 Top-3 Top-4 Top-5

Recall rate (%) Our method

(similar images) 70 77.67 84.33 88.67 91.67 51.4

Our method

(similar species) 65 82 88.33 93 94.33 51.4

Lee-Chen’s method 82.33 - - - - 48.2

In order to compare the performance with Saitoh-Kaneko’s [1] method, we

collected the same numbers of species and images with Saitoh-Kaneko’s database, the

reason is that we can not get their database. The recognition results are listed in Table

7. We can see that the recognition rate of our method is much higher than

Saitoh-Kaneko’s method.

Table 7 Performance comparison between our method and Saitoh-Kaneko’s method

using the same numbers of Saitoh-Kaneko’s database.

Recognition rate (%)

Top-1 Top-2 Top-3

Our method 97.5 100 100

Saitoh-Kaneko’s method 96.03 99.26 99.26

Finally, we compare the performance of our recognition methods: flower, leaf

Only flower images of 24 species; (ii) Only leaf images of 24 species which

correspond to (i); (iii) A pair of flower and leaf images of 24 species. Table 8 shows

the performance results. From Table 8, we can see that the combining method get

higher recognition rate than those using only flower or leaf image. Hence, the

combining recognition method is more effective and can provide better results to user.

Table 8 Performance comparison.

Recognition rate (%) Method

Top-1 Top-2 Top-3 Top-4 Top-5

Number of images

Number of species Our method

(Flower) 76.9 93.1 98.3 99.1 99.7 684 24

Our method

(Leaf) 68.5 90.4 96 98.3 99.4 480 24

Our method

(Combining) 94.4 99.7 100 100 100 684 24

We have built a plant recognition system written in Java language on a PC. Figs.

22(a), 22(b) and 22(c) are the interfaces for the flower, leaf and combining

recognition systems, respectively. Figs. 23(a), 23(b) and 23(d) are the interface for

recognition results of flower, leaf and combining recognition systems respectively.

After we retrieved the candidate images, users can click on the image to get the

system to know the species of the plant which they did not know before.

(a) (b)

(c)

Fig. 22 Interfaces of recognition system. (a) Flower recognition system. (b) Leaf recognition system. (c) Combining recognition system.

(a)

(b)

Fig. 23 Interfaces for recognition results. (a) Flower recognition results. (b) Leaf

(c)

(d)

Fig. 23 Interfaces for recognition results. (a) Flower recognition results. (b) Leaf recognition results. (c) The retrieved information of the query image. (d) Combining recognition results.

CHAPTER 5 CONCLUSION

In this thesis, we have proposed a plant recognition system based on leaf and

flower. In the flower recognition system, we use an automatic segmentation based on

human visual system. Then, a simple and interactive user interface is applied.

According to the shape and color features of the flower, 14 features are extracted from

the segmented flower image. Finally, a similarity measure is provided to do

recognition.

In the leaf recognition system, we also proposed an automatic segmentation

method and a solution to treat rotation problem. Next, we extract 5 features according

to the characteristics of the leaf shapes. Then, we preserve possible species and find

out the similar images from leaf image database by similarity distance.

In the combining recognition system, a new method has been proposed for

recognizing plants based on leaf and flower. Firstly, the features of the leaf and flower

are extracted. Next, we calculate the similarity between the query image and database

images of leaf and flower and then combine the results of leaf and flower. Finally, the

system can find out the most similar species.

recognition system. This means that our combining recognition system can get higher

accuracy rate than the single recognition system.

REFERENCES

[1] T. Saitoh and T. Kaneko, “Automatic Recognition of Wild Flowers,” Proc.

International Conference on Pattern Recognition, Vol. 2, pp. 507-510, 2000.

[2] T. Saitoh and T. Kaneko, “Automatic Recognition of Wild Flowers,” Systems and

Computers in Japan, Vol. 34, pp. 90-101, 2003.

[3] J. B. MacQueen, “Some methods for classification and analysis of multivariate

observations,” Proceedings of 5-th Berkeley Symposium on Mathematical

Statistics and Probability, Berkeley, University of California Press, 1: pp.

281-297, 1967.

[4] L. J. Chen, “A Fast Automatic Segmentation Algorithm based on Local

Color-Distribution,” Master Thesis, Institute of Multimedia and Engineering,

National Chiao Tung University, Taiwan, ROC, 2007.

[5] Dan Margulis, “Photoshop Lab Color: The Canyon Conundrum and Other

Adventures in the Most Powerful Colorspace,” ISBN 0321356780.

[6] Lab color space, Wikipedia Google, URL:

http://en.wikipedia.org/wiki/Lab_color_space.

[7] J. Zou and G. Nagy, “Evaluation of Model-Based Interactive Flower Recognition,”

[8] T. Saitoh, K. Aoki and T. Kaneko, “Automatic Recognition of Blooming Flowers,”

Proc. International Conference on Pattern Recognition, Vol. 1, pp. 27-30, 2004.

[9] E-N. Mortensen and W-A. Barrett, “Intelligent Scissors for Image Composition,”

Proc. Computer Graphics and Interactive Techniques, pp. 191-198, 1995.

[10] Z. Wang, Z. Chi and D. Feng, “Shape Based Leaf Image Retrieval,” IEE

Proc.-Vis. Image Signal Proc., Vol.150, No. 1, pp. 34-43, 2003.

[11] C.L. Lee and S.Y. Chen, “Classification of Leaf Images,” International Journal

of Imaging Systems and Technology, Vol. 16, No. 1, pp. 15-23, Jul. 5, 2006.

[12] N. Otsu, “A Threshold Selection Method from Gray-level Histogram,” IEEE

Transactions on Systems, Man, Cybernetics 8, 1978.

[13] T. H. Hsu, “An Interactive Method for Flower Recognition,” Master Thesis,

Institute of Multimedia and Engineering, National Chiao Tung University,

Taiwan, ROC, 2008.

[14] HSL and HSV, Wikipedia Google, URL:

http://en.wikipedia.org/wiki/HSI_color_space.

[15] H. Y. Huang, “An Automatic Recognition System of Leaves,” Master Thesis,

Institute of Multimedia and Engineering, National Chiao Tung University,

Taiwan, ROC, 2008.

[16] Plant taxonomy, Wikipedia Google, URL:

http://en.wikipedia.org/wiki/Category:Plant_taxonomy.

[17] http://www.ecse.rpi.edu/doclab/flowers.

相關文件