| Classification and identification of plants is a basic important work with long-term. It can bring people better understanding and using of plant resources. This work has a key role in distinguishing plant species, exploring genetic relationship between them, and clarifying the evolution law of plant system. Professional experience is used to identify plant species through their external morphological features in traditional methods mostly. These methods are difficult for non-professionals to grasp or to be applied in daily life. On the other hand, there are about 250,000 kinds of flowering plants has been named and classified on the earth. So it's hard to identify them effectively just through memory or experience.Prunus mume was studied as research object in this paper. Image segmentation, feature extraction, classification, etc. were studied based on prunus mume's image. A prunus mume species recognition system automatically was established based on these technologies. The key work and contributions in this paper are as follows:1. Plant species identification technologies including image segmentation, feature extraction, and classification category based on image analysis were analyzed and researched deeply. On feature extraction, the base visual feature extraction and local feature extraction methods were analyzed respectively.2. A segmentation method for prunus mume's image with natural background was proposed. The method is based on texture feature and color feature segmentation (T&C, Texture feature and Color feature). The prunus mume flower image's characteristics was analyzed in this algorithm. For prunus mume's leaves come after flower, the branches and overlap flowers are the main interferences in the background without no large leaves gathering. It can be segmented by a method combining fractal texture feature and color feature. Fractal feature was used for the texture segmentation. Results show that the edge of flower and branch has the highest fractal dimension, and most other interferences can be cleared in this step. The remaining interferences can be segmented by color features. This paper improved color histogram cumulative algorithm. The branches and flowers were segmented from two perspectives respectively:hue and saturation. Finally target area was segmented through subsequent processing. Through comparison with 2RGB color model segmentation method and Crab Cut algorithm in the experiments, T&C segmentation algorithm is proved to be quite effective for prunus mume image segmentation. The 2RGB model is not suitable for this kind of data. Much interaction and time-consuming are existed in Grab Cut algorithm.3. The algorithms for color, shape, and texture feature description respectively about prunus mume were proposed, and some of them were improved. The visual feature and local feature were extracted of the target region. Color features, texture features and shape features were used in visual feature. In color feature extraction, some characteristics of prunus mume images were analyzed. The results of histogram classification present the method suitable for describing the color characteristics of prunus mume image. Convex hull area ratio, flatness, and border sequence matrix were used for describing shape features. When calculating the flatness, the method calculating the original block was improved. Because the image has been segmented, the background is composed of black pixels. Only the flower region is useful to be calculated for flatness. The algorithm is simpler than original method and the final results were not be affected using the improved method. For the texture features, the fractal dimension and the gray level concurrence matrix(GLCM) were used for prunus mume image. The flower's smallest external rectangular region was extracted firstly as the new length and width of the image before calculating the GLCM. The GLCM of 00,450,900 and 1350 were calculated. Then the corresponding matrix elements were cumulated and divide the results by 4. The average matrix was used for the new GLCM, and the related parameters of the new matrix were calculated as texture feature description. Including color feature, shape features and texture features, the visual features are a 19-dimensional feature vector. Local feature extraction also was described in this paper. Finally the characteristics and principal component of visual features were analyzed. The results show that the 19th characteristic plays a minor role, and the other visual features extracted in this paper all have their own role for classification with little redundancy.4. This paper improves traditional three-layer BP neural network classifier. A hidden layer was added and related parameters were improved. In the experimental section, SVM classifier and BP neural network classifier were used separately. The results show that that traditional BP neural network's classification results are not satisfactory. After adding a hidden layer, the recognition ratio has improved significantly. A lot of analysis on the choice of SVM kernel function's parameters were made and the best parameters were obtained. A large number of experimental analyses show that the recognition system is suitable for prunus mume species identification. The proposed algorithm can also be suitable for ordinary flower image recognition with flexible portability through experiment.
|
PDF Full Text Request