Study On Multi-cue Plant Species Identification Method

Plant recognition is the automatic recognition of plant names from a given image.Most of the existing research on automatic plant recognition focuses on the identification of plant species using a single organ,such as flowers,leaves or fruits,or using a single image to identify plants.However,it is not reliable to identify plants by using individual organs of plants,because many different plants have very similar organs.For the pictures directly collected in the field,there are usually complex backgrounds,so it is difficult to extract the most distinguishing features of plants from the pictures by using a single picture.In order to overcome the low accuracy of plant image recognition,this paper studies the multi-clue plant species recognition method based on image set and proposes a multi-clue plant recognition method based on depth convolution neural network.The main contents and conclusions of this paper are as follows:(1)Two imageset-based algorithms are improved and applied to plant recognition.Two classical image set algorithms are improved.Based on pairwise linear regression image set classification and prototype discriminant learning image set classification,the idea,algorithm principle and algorithm flow chart of the improved image set algorithm are introduced in detail.The improved algorithm is applied to plant recognition,and plant recognition model is trained with plant image set.Because the background of plant image is complex and similar to the plant to be identified,in order to better highlight the plant characteristics,data enhancement technology is used to expand the number of plant images.And the accuracy of the last two improved plant recognition models based on image set is 62.01% and 64.22%.(2)Improve the network structure of GoogLeNet and apply it to plant recognition.New Inception module structure is proposed,and a new activation function h-Swish is used to replace the traditional activation function ReLU.Twelve types of plants were selected from the PlantCLEF 2016 data set and used to train a single organ model based on the improved GoogLeNet.By using transfer learning with ImageNet and adjusting the number of neurons in the final connective layer of the improved GoogLeNet,the optimal single organ classifier for flowers,fruits,leaves and plants is trained.Finally,the recognition accuracy of single classifier model for flower,fruit,leaf and whole plant is 82.74%,86.23%,63.25% and 65.06%.(3)A multi-clue plant species recognition method based on improved GoogLeNet network is proposed.According to the predicted labels and scores of each single classifier,multi-organ fusion recognition is proposed.Each single classifier has different weights according to its average recognition accuracy.The final predicted categories of plants are determined by the predicted labels of the single classifier,the predicted scores given by the single classifier and the weights of the single classifier itself.In the PlantCLEF2016 data set,the multi-cue model proposed in this paper has a recognition accuracy of 92.33% for twelve kinds of plants.The experimental results show that the recognition accuracy of plant images with complex background based on traditional image set recognition method is not high,and more elaborate preprocessing is needed to achieve higher accuracy.The recognition accuracy of multi-cue model based on deep convolution neural network is significantly higher than that of plant species recognition method based on image set.