Application Of Convolutional Neural Network In Agricultural Scene

Agricultural image is a kind of data acquired from agricultural scene by image acquisition system,which carries a large amount of information reflecting crop growth or other specific interest.Automatic detection of agricultural image and automatic identification of regions of interest are helpful for managers or automated equipment to make efficient decisions in the process of agricultural production.Convolutional Neural Network(CNN)is the fastest growing and most widely used computer vision algorithm in the field of international academia and industry in recent years.It is the frontier technology and research hotspot in the field of computer vision.It is also an effective tool to solve practical engineering problems.In order to improve the intelligence of agricultural production process,the latest research results of computer vision are combined with agricultural images,and the agricultural images are analyzed by means of advanced convolution neural network technology.This paper summarizes the methods and development trend of agricultural image analysis by using image processing and computer vision technology by domestic and foreign scholars.The development process and theoretical basis of convolutional neural network algorithm are systematically reviewed.Aiming at the problems of disease classification,plant organ recognition and location,and weed segmentation in agricultural images,an agricultural image analysis framework based on convolution neural network was proposed.The transfer learning was applied to Alexnet convolution neural network to classify 10 kinds of tomato leaves from diseased and healthy leaves.Using 14529 tomato leaf disease images,70% of them were randomly selected as training set and 30% as validation set.The structure of Alexnet convolution neural network model was migrated.The mature Alexnet model and its parameters were trained on Imagenet image data set to identify tomato leaf disease.During the training process,fixed low-level network parameters unchanged,fine-tuned high-level network parameters,tomato disease images were input into the high-level parameters of the training network,and the trained model was used to classify 10 kinds of tomato leaves,and 20 groups of experiments were carried out.The results show that the disease classification model based on transfer learning can classify 10 kinds of Tomato Leaf Diseases quickly and accurately.Aiming at the requirement of vision system of agricultural robot,the image data before cotton picking is taken as the research object.The SSD(Single Shot MultiBox Detector)method and the plant organ recognition and spatial location method of depth camera are proposed.This method is applied to cotton top for target recognition and spatial location.Firstly,a convolution neural network model based on SSD method is established to locate the cotton apex,and then the spatial coordinates of the cotton apex are calculated by combining the depth camera data and camera internal parameters.An experimental platform was built to verify the proposed method.The experimental results show that verifies that better recognition results can be obtained under different illumination conditions,and the model has reliable recognition ability for targets of different scales.Finally,the feasibility of this method is verified by practical experiments.In order to improve the recognition accuracy and real-time performance of crops and weeds,a pixel-by-pixel classification method for real-time agricultural images based on depth separable convolution was proposed,taking the field color images of Sugarbeet Seedlings as the research object.In this study,the field color images of Sugarbeet seedling collected by agricultural robots were used to label each pixel in the color image into three categories: crop,weed and soil by manual pixel-by-pixel labeling method,and the labeling information of a single category was placed in three different image channels to form a data set for training and testing.Firstly,a deep separable convolution neural network model based on coder-decoder is established,which combines the coder part with the decoder part in multi-scale.The encoder part determines the location of the pixels,and the decoder part obtains the classification of the pixels.Then,in order to solve the problem of unbalanced coverage of the classification category,the low coverage rate is improved through single channel annotation information training.In order to control the scale of network parameters,the number of convolution kernels of control points is multiplied by width.At the same time,the network model is further tested under different resolution input conditions to discuss the real-time performance of the network model.