Research And Implementation Of Plant Leaf Recognition Algorithm Based On Deep Network Feature Learning

Plants are an essential part of our human existence.It is important for us to know plants and understand plant characteristics.With the advent of the era of big data and the development of digital media technology,we can access to a large number of plant images conveniently.Facing such vast amounts of data in the network,it is the big challenge to quickly and accurately identify plant species for understanding the characteristics of current field of plant leaf recognition.In recent years,leaf recognition has been studied in the field of computer vision.Researchers focus on how to improve the accuracy and time efficiency of the recognition.With the development of neural network science and artificial intelligence,people get rid of the single learning characteristics,through the integration of various disciplines,so as to achieve a better recognition effect.In this paper,the combination of feature learning,machine learning and neural science is applied to leaf recognition.Learning is the core of intelligent recognition.Therefore,this paper mainly focuses on the deep research for leaf recognition algorithm which is based on the deep network study of the characteristics.The main research contents include the following three aspects:First of all,our previous research is mainly concentrated on texture feature extraction.For a set of feature vectors,we use machine learning method to as a classifier for training and predicting.To ensure the leaf rotation invariance,we use Binary Gabor Pattern algorithm(BGP)to extract detail texture features,and choose a more extreme simple and effective learning machine Extreme Learning Machine(ELM)as the classifier.Then,we found that for our gradually expanding database,the early methods,whether in accuracy or time,are not particularly satisfactory in performance,but also in the aspect of feature extraction,more manual intervention.Therefore,we mainly use the framework of deep learning in the later research.We use convolutional neural network(CNN)to design the network structure,adjust the parameters of the network,and autonomously study the deep features of learning layer for the training set and test set.And get the feature model to predict.Moreover,we use the deconvnet to visualize each layer and display the characteristics of learning,in order to verify the rationality and effectiveness of the network.Finally,according to the two previous study programs,we designed two sets of leaf recognition systems to verify the feasibility and rationality.For previous system,displaying feature histograms and similar top 10 leaves(including similarity values and categories)in the database for any input leaves.For later system,visualization of any test image feature maps(9 maximal activation maps per layer)and the top 5 similar leaves(including accuracy and category)in the database.At the same time,from the recognition effect,analysis and comparison test results of the former and latter methods.