Research On Recognition Method Of Tomato Leaf Deficiency Based On Convolution Neural Network

Tomato is one of the most widely cultivated vegetables in the world,and plays a key role in the supply of vegetables in China.The growth of tomato requires the fertility supply to meet both high consumption and high supplement conditions.There is blindness in the application of fertilizer in the actual production process,and it is easy to appear the symptoms of soil obstacle and deficiency of element,if found not in time will cause huge losses in production.This paper aims at the problem that the image features of tomato leaves are little changed and it is difficult to capture the detail features,based on the RGB image data and hyperspectral image data,the classification model of tomato deficiency image was designed after symptom and before symptom,which provided a reference for high precision intelligent recognition of tomato deficiency.The main work of this paper is as follows:(1)A multi-source data set based on RGB and hyperspectral images is constructed.The growth environment factors of tomato seedlings were regulated by the artificial climate chamber,and the experiments of tomato nutrient deficiency culture were carried out.A normal control group and three nutrient deficiency groups were established in the experiment,which were nitrogen,phosphorus and potassium deficiency groups.A total of 6250 images of tomato leaves containing nitrogen,phosphorus and potassium deficiency and healthy tomato leaves were constructed.A total of 327 hyperspectral image data sets of tomato leaves containing nitrogen,phosphorus and potassium deficiency and healthy tomato leaves were constructed.(2)An image classification method based on attention mechanism and multi-scale feature fusion is proposed.In order to solve the problems such as the difference in size of characteristic region,the difficulty of extraction and the difficulty of distinguishing,which are caused by the inconspicuous deficiency of tomato and the great difference in the characteristics of each growing period,an image classification method based on attention mechanism and multi-scale feature fusion convolutional neural network-msff-am-cnns was proposed.Firstly,a feature fusion Module(MSFF Module),which is composed of multi-scale convolution kernel,is proposed to solve the problem that the perceptual effect of fixed-scale convolution kernel is not obvious for different size of tomato,msf-am module and Deep-MSFF Block are improved to extract shallow features,and multi-level global information and multi-feature channels are combined to selectively highlight information features and achieve accurate feature location Finally,Focal Loss is introduced to focus the training on complex and difficult-to-classify samples.The results show that the MSFF-AM-CNNs model can meet the needs of highprecision classification of tomato leaf images,and it has a wide range of application.(3)A method for classification of hyperspectral images of early-stage nutrient-deficient tomato leaves was proposed.For the Classification Task,a hyperspectral image pre-processing algorithm based on MSC is designed to remove noise and irrelevant variables,and to screen out abnormal samples before Data pre-processing,at the same time,low variance filtering feature selection is carried out to eliminate the features which have no effect on classification,so as to improve the efficiency of the following key features extraction CARS,IVISSA and SPA algorithms are used to extract the key wavelength of tomato deficiency feature,and the best feature extraction method is selected according to the results The hyperspectral data structure is changed by transforming the one-dimensional spectral sequence into the twodimensional spectral data and fusing the global spectral information,a hyperspectral image classification model based on two dimensional convolutional neural network was designed for early stage nutrient deficient tomato leaves.Experimental results show that the Precision,Recall,F1 score and Accuracy of the model are significantly improved,and Accuracy can reach 98.06%,the results show that this method has strong applicability in the task of tomato leaf image classification.