Research On Methods For Crop Lesion Image Processing And Disease Recognition With Machine Learning

Agriculture, which is a complex system to center on organism, always is ecologically and physiologically complex in respect of production processe and related closely to territoriality. The Internet of Things(Io Ts) is an important engine to drive the reformation of agricultural science and technology, and it is selfevident to accelerate the development of agricultural Io Ts information technology becomes the common choice of all countries and regions. China, which is a country of agriculture power in the world, makes huge demands on agricultural information technology. At present, most of the agricultural equipments or terminals can only acquire crop image, and they are incapable of extracting some useful information to guide agricultural production and management because of lack of feature extraction. Automatically analyzing, recognizing, processing the image from the network terminal information, and enabling the system to response to the growing information as an intelligent life has become the primary task to get rid of the dilemma of the "blind" to vision sensible equipments. Image of growing crops intuitively and lively represents growth, development, health-damage details, lesion causes and other aspects of the information related to crops. Developing the function of machine vision equipment by which to approximately correctly read the agronomy information contained in image, then to provide farmers a scientifc and realtime instruction for agriculture jobs has become the primary goal of the development of smart agriculture.In the dissertation, an in-depth research is carried out on the penalty regularization method for SVN(support-vector network) based on sparse training set and its application to seed quality evaluatine with CT image, preprocessing of crop lesion image for disease recognition, reducing image dimensionally with numerical analysis and computaion for lesion recognition, dimensionality reduction of lesion image based on semi-supervised deep learning of RBM network, the incorporating network of dimensional reducion and pattern recognition with deep convolutional learning.It extends the theory and application of machine learning, especially of deep machine learning, and brings forward the data structure and algorithm paradigm of intelligent analysis & processing of agricultural information frow the view of machine learning.Finally, the proposed theory and methods are applied in the orchard production scene like a factory of fruit for apple lesion image recognition and early disease-pest warning, and are verified in respect to algorithm by experimentation. The work provides valuable reference for the study of crop lesion image recognition with agricultural intelligent system based on deep machine learning. The main contribution of this paper is as follows.1.Method of evaluating and grading seed quality based CT image with SVN.It is discussed in depth where an unbalanced, sparse training set is used to train a SVN, the outputted classifiers whose mis-prediction rates are badly imbalanced, are always output and badly unavailable. Based on analysis of Lagrange multiplier, the SVN learning algorithm with penaltyregularization and the method for penalty coefficientregularization are proposed. The algorithm is applied to grading wheat seeds quality by a feature data set of CT image, and experiment result demonstratesclassification accuracy of concerned sparse object is improved evidently; a very good global performance is presented with an amusing availibity and generalization.2. Preprocess of lesion image sampled in a complex scenario.In contrast with the sampling of the widely used bentchmark datasets, working conditions for sensors installed in the farm-orchard are complex, with several interference, and both the mobileand the immoblie photographing means is difficult to assuredly sample a lesion image dataset where the representative image of each state distributes uniformly under constraint of some restricted conditions and limited sampling times. Before joining into learning and test,lesion image requires a preprocess procedure where noise is eliminated and size is formatted.Take apple for example,lesion image sampling and lesion recognition is formulated and preprocess procedure of lesion image and several involved methods are proposed such as reshaping, emulation of orientation and brightness diversity,sparce coding.3. Dimensional reduction of image for lesion recognition with numerical analysis and computation. Around an analysis of mathmatical principle of PCA, PCA dimensional reduction of discarding overflow eigenvalues is proposed,which presents better reconstruction performance using "95" norm. Somegeneralization experiment are done upon ORL face dataset, and 3D accuracy surface suggests this method also is effective to extract face feature on condition that training prarametes are selected opitimally and rationally. In the second stage of this research, combining SVN and PCA dimensional reduction as a means of feature extraction, the method of lesion recognition with image is constructed, which shows an amusing and promising experimental performance. Dimensional reduction of image for lesion recognition with semi-supervised deep learning network.On the analysis of mathimatical formulation of restricted boltzman machine(RBM) energy model, the method of lesion image dimensionality reduction with semi-supervised deep learning is brought forward,of which reconstruction performance is better improved than PCA. Aimed at pretraining of the RBM network, the method named by "Random step contrastive divergence based on feedback" is pioneered around k step contrastive divergence(k CD). Comparison experiment between "Random step contrastive divergence based on feedback" and kcd demonstrates the proposal shows a longer convergent time, but after the object funtion converges it present better performance stability, an optimal training object function value, and it is more benifical to system perfomance stabilization.4. Incorporated supervised deep learning network for dimensionally reduction & recognition of lesion image. As to the subject, the method for pathologic image recognition-diagnosis is proposed based on a deep learning convolution neural network, whose toplogic structure and error back-propagation appropriative for convolution is also devised. In addition, an innovative method for updating free parameters so-called the gradient descendent with flexible momentum is brought forward and its two embodiment is designed,i.e. linear varibale momentum and quadratic varibale momentum. Using an incorporate network to integrate feature extracion and pattern recognition, and achieving a collaborative training of recognition parts and extraction parts with a sharable learning machanism the method resolves the mismatch of course-goal and its converging graph demonstrates its relatively fast convergence speed. In parallel with traditional momentum gradient descedent and self-adptive momentum, the convergenceepoch of both linear momentum and quadratic momentum presents a large advancement.