Research On Plant Leaf Recognition Algorithm Based On Repeated Image Segmentation And Complex Frequency Domain Features

The recognition of plant species has always been one of the most challenging areas of computer vision.Accelerating the development of plant recognition technology can effectively protect plant diversity,maintain ecological balance and promote social development.Considering that the leaf has the advantages of collection and storage,the leaf is used as the research object of plant recognition.There are four main stages in plant leaf recognition,including image acquisition,image preprocessing,feature extraction and image recognition.During acquiring the image,the machine is susceptible to illumination and shooting angles which result in uneven illumination and shadowing on the leaf image,which will have a certain impact on the subsequent segmentation and recognition on the leaf image.And now researchers mostly extract features in the spatial domain while extracting features in the spatial domain often introduces more redundant information to reduce the effectiveness of features.So it is important to study how to improve the image segmentation accuracy and reduce the redundant information in features for plant leaf recognition.This paper has studied the image segmentation and feature extraction for plant leaf images in Leafsnap and Flavia datasets.The main contents of this paper are as follows:(1)This paper proposed a leaf image processing algorithm based on repeated image segmentation.Aiming at the problem that the image was difficult to segment because of uneven illumination and shadow,the algorithm first converted plant leaf images from RGB space to HSV space which was adapted to human eyes.Then,the V channel in the HSV space was taken out,and a homomorphic filtering process was performed on the V channel.This could perform illumination compensation on the image without changing the color information.In order to protect the leaf region as much as possible when performing image segmentation.,the k-means clustering segmentation algorithm was used to remove the obvious background area firstly,and then the bimodal minimum segmentation algorithm was used to find the boundary value between the leaf area and the shadow area,and then the second segmentation was performed.Finally,the petiole and the remaining noise blocks was removed by a morphological method.(2)This paper proposed a plant leaf recognition algorithm based on complex frequency domain texture features.Aiming at the problem that extracting features in the spatial domain would introduce more redundant information and reduce the validity of features,the algorithm performs block processing on the processed leaf image firstly,and the segmentation helped to preserve the local detail information.Then,each image block was decomposed in the complex frequency domain by performing double-tree complex wavelet transform and the decomposed 6 high frequency sub-bands and 2 low frequency sub-bands were extracted.Next,the local binary pattern feature was calculated for high frequency sub-bands and the local phase quantization feature was calculated for low frequency sub-bands,and the two features were combined to obtain the feature of the image block.In order to reduce the interference of valid information such as background,a weight factor was introduced to weight each image block feature.Next,the features of all image blocks were concatenated to obtain the feature of the entire image,that is,the complex frequency domain texture features.Finally,all training features and test features were calculated and were classified by KNN or SVM classifier.(3)This paper performed comparative experiments about repeated image segmentation and complex frequency domain texture feature recognition algorithm on two datasets.In this paper,Flavia and Leafsnap were selected as the dataset of plant leaf with high comparability,and the leaf image segmentation experiment and the leaf image recognition experiment were carried out respectively.The experimental results of leaf image segmentation showed that the repeated image segmentation algorithm had better segmentation effect than other image segmentation algorithms.The repeated image segmentation algorithm has achieved good results in the three evaluation indexes of segmentation error,false positive rate and false negative rate where the average values were 7.40%,1.50%,and 0.86%on Leafsnap dataset respectively.The experimental results of leaf image recognition showed that the complex frequency domain texture features were more descriptive and better than other features.And our algorithm had certain advantages compared with other literature algorithms where the highest recognition rate on the Flavia dataset was over 95%.