Local Rose Breeds Detection System Using Transfer Learning Techniques

Bangladesh is known as the land of flowers. At recent, the flower cultivation covers 10,000 hectares of land. The number of farmers, who are growing flowers is more than 5,000 and around 150,000 people are engaged in flower business for their subsistence [1]. From the ancient time, people are using roses for many purposes. Roses are the most desirable flowers for any kind of occasion. Rose water is the a very valuable thing that people are using according to their needs. Many perfumes have been made of rose oil. Besides making sweets, rose petals are being used in potpourri. Rose-hip is a fruit, which is abundant with vitamin c. In Bangladesh flower business in shortened without roses. The diversity of roses is improving day by day. Rosaceae is a plant family enlarged with over hundred genera and more thousands of species. As roses play a vital in flower business so knowing about all breeds of roses is really important. In Bangladesh, nine types of rose breeds have seen so far. These nine breeds of roses are Papa Meiland, Iceberg, Rose Gaujard, Bengali, Sunsilk, Queen Elizabeth, Julia’s Rose, Dutch Gold, King’s Ransom [2]. Not only from outer look, these rose breeds are different from inner part and that’s why they need different care for cultivation. Breed detection will be very helpful in the sector of rose cultivation and flowers business. As roses are the most common and demanded flower so people who cultivate flowers give more attention on cultivation of roses. According to the demand and the amount of rose cultivation, breeds detection of roses will be very beneficial. By knowing the breeds, it will be easier to detect the cultivating process and also the proper way of taking care of rose plants and flowers. All the rose breeds which are presently available in Bangladesh are basically hybrid breeds, so by detecting the rose breeds and comparing the pros and cons of all the breeds it will be possible for flower analysts to create or invent new rose breeds according to the climate of Bangladesh. Rose breed detection will play a vital role in rose cultivation. If cultivators will be aware about rose breeds, they will be educated about the cultivation of each breed. They will act based on the characteristics of the breeds, like the amount of fertilizer they need, how many days they take to grow, the limitation of a breed, caring method and many more. The two main goals of this research are as follows:

• •

  Develop an expert breed detection framework that can process the images of rose breeds and recognize the category of a particular breed.

• •

  Maximized the data set amount by applying some image augmentation techniques, background noise removal, and data simulation.

The rest of the portion of this research paper is organized as follows. Section II summarized the important literature reviews. Proposed methodology along with data set distribution are described in section III. Experimental result evaluated and discussed in the Section IV. Finally we have concluded our work and narrated the future scope in section V.

As per study on previous work, there is no work that has been presented on breed detection of a particular flower. There is various work on flower classification and identification of flower species. Some literature studies are below:

Philipe A. Dias et al.[3] have proposed a CNN based apple flower detection system. They have collected a total 147 images apple trees. To perform feature extraction, they have used fully connected layer of CNN and finally classified the images using SVM. Their CNN+SVM model achieved more than 90% in terms of recall and precision score.

Xiaoling Xia et al.[4] classified flowers using the transfer learning technique based on Inception-v3 model. To classify flowers, they have used one data-set contains 17 species of flowers and for each species they have collected 80 flower images and another one contains 102 species of flowers and they have collected 40-258 flower images for each species. Their model obtained the classification accuracy 95% on first data-set and on second data-set they have got 94%.

I.Gogul et al.[5] proposed a system to recognize flower species using CNN. They have collected two datasets. For first dataset, they have collected 2240 images of flowers from 28 categories and for second dataset, they have collected 8189 images of flowers from 102 categories. They have used transfer learning approach for feature extraction. The highest accuracy of their model was 93.41%.

Hazem Hiary et al. [6] represented two step deep learning classifier to identify flowers. They have modeled a binary classifier in a fully convolutional neural network framework and sturdy CNN classifier to identify the different flower types. They have used three datasets. The datasets have 8189 images from 102 categories, 1360 images from 17 categories and 612 images from 102 categories flowers respectively. They have got the classification outcome more than 97% on every dataset.

M. Cıbuk et al. [7] used a hybrid method based on DCNN. For features extraction, they have combined the features from AlexNet and VGG16 models. At the end, they have applied mRMR method to specify more effective features. To identify the flower species by using the extracted features, they have applied the SVM classifier with RBF kernel. They have used two datasets of 17 and 102 flower categories accordingly. They obtained the accuracy rate 96.39% and 95.70% for the datasets individually.

Thi Thanh Nhan Nguyen et al.[8] tshowed the sturdiness of DCNN to identify flower species. They have extracted a flower dataset of 967 species from PlantCLEF 2015. The accuracy exceeded 90% for that dataset.

Yuanyuan Liu et al.[9] prepared a method based on CNN for classification of flowers. They have created a dataset of 52,775 images from 79 species of flowers. By using this dataset, they have got 76.54% of accuracy and they applied their method on a well-known dataset and got 84.02% of accuracy.

Saiful Islam et al.[10] have developed a system to classify local flowers using CNN. Their dataset has 6400 images which belongs to eight kind of flower species. Their system has scored 85% classification accuracy.

Busra Rumeysa Mete et al.[11] have classified flowers by using Deep CNN and machine learning approaches. They have used Deep CNN to perform feature extraction and to get better outcome they have showed the utilization of image augmentation. Finally, they have compared some machine learning approaches. They have used two datasets. They have achieved the highest accuracy 98.5% and 99.8% by using SVM and MLP classifier respectively.

Yong Wu et al.[12] have built a method to recognize flowers using CNN and transfer learning technic. For feature extraction, they applied DCNN. They have combined neural network model with transfer learning technic. By achieving a better accuracy of flower classification than other traditional methods, they have showed a remarkable enhancement.

S. M. Farhan Al Haque et al.[13] proposed a system for the detection of guava disease using deep learning models. Their final model achieved 95.61% of accuracy.

The analysis of previous research works found that various researchers have applied different techniques to detect different fruits and vegetables. But nobody worked on flower classification and identification of flower species. We aim to predict rose breeds that can help the farmers and agriculturists to maximize the production of different roses.

In this portion, we have found the model, which performed better for the dataset that we used to detect selected rose breeds. The presented working methodology has shown in “Fig. 1”.

In this research, to train the models, we used 1550 images of five classes and to test the model, we used 387 images. We compared the confusion matrix of different algorithms to evaluate the accuracy. In this study, we have used four models. These are Inception V3, Xception, ResNet50 and VGG16. These are the most popular and used models in image identification and classification. To get a comprehensive knowledge, we have researched and compared the outcome of each transfer learning models. For each model, we have tried 20 epochs. To evaluate the performance of the models we have used confusion matrix. As there are a total five classes in our dataset, we used 5x5 matrix to calculate. “Fig. 4”, “Fig. 5”, “Fig. 6”, and “Fig. 7” shows the confusion matrix for the Inception V3, Xception, Resnet 50 and VGG16 models respectively.

“Table. II” shows the accuracy and loss comparison for all the four models that we have applied. According to the “Table. II” it is shows that VGG16 outperformed in terms of accuracy. This model shows the highest accuracy among the other three models. The ROC curve for the VGG16 model is shown in “Fig. 8” .

The accuracy of model VGG16 and the model loss are shown in “Fig. 9” and “Fig. 10” respectively. In “Table. III”, we have shown a comparative analysis of our work with some other researches regarding flower breed detection.

In this research work, we have proposed and discussed about a significant Algorithm for breed detection. In addition, we have used a real-world dataset which contains 19,39 images of five different breeds of rose. we have applied four Transfer Learning methods, these are Inception V3, VGG16, ResNet50 and Xception. Among these four methods, the performance of VGG16 was great and achieved the highest accuracy of 99%. We would like to build a Transfer Learning methods-based system in future. This system will not only detect the breed of a rose but also it will provide the characteristics of that particular breed and the cultivation and caring process of the breed. “Fig. 11” shows the proposed web-based architecture of rose breed detection.

In this system, at first a user needs to capture a picture of a rose or upload a picture. Then, when the picture will be sent to the system, the system will work on it. It will analysis the picture and it will compare the breed with the breeds which are present in the database. If it gets any similarity in between the input data and the data of the database, it will display an output with the name of the breed of the rose with the details of that breed and also will provide the proper way of cultivating and taking care of the breed. As we have used transfer learning techniques instead of traditional CNN, the system is complex and more time consuming.