Research On Seed Intelligent Sorting Algorithm Based On Convolutional Neural Network

How to identify and classify different types of seeds and abnormal seeds efficiently and accurately is still a relatively difficult problem in the field of agriculture.The traditional artificial detection and image processing algorithm has the problems of low efficiency and low adaptability in crop seed sorting,so it is of great significance and application value to find a method that can replace the traditional seed sorting.As one of the basic deep learning algorithms,convolutional neural network has the characteristics of simple structure and strong adaptability.However,compared with the traditional image processing methods,the convolutional neural network algorithm has a huge amount of parameters and computation,which seriously hinders the application in the industrial equipment of the embedded system with limited computing resources.In order to study the image processing and pattern recognition of seed classification,this paper uses convolution neural network algorithm of deep learning technology to recognize and analyze the sunflower seed image and corn seed image.1)Based on the current advanced deep learning algorithm,a reduced convolutional neural network model for seed sorting task is proposed.Firstly,alexnet,VGg,RESNET and other networks were used to verify and compare the sunflower seed datasets.Then,the convolution layers of resnet18 are visually adjusted,and the parameters such as the size and number of convolution kernels are constructed.The resnet10 network can significantly reduce the amount of parameters and computation while retaining the recognition accuracy of the original network.2)For convolutional neural network model still has many redundant parameters for seed image data with fewer learning categories,lhcnet is constructed by introducing heteronuclear convolution and channel attention mechanism module on resnet10.First of all,the complexity of seed image is far less than that of natural image.Therefore,the fusion of 1 × 1 convolution and 3× 3 convolution with small size and less computation can effectively reduce the computation and parameter of model and improve the reasoning speed of image recognition without reducing the ability of network to extract seed image features.Secondly,the effective seed image feature information extracted by the partial convolution kernel in the convolution layer is very little.Therefore,after the convolution layer,the channel attention mechanism module is introduced to learn the importance of each channel feature,so as to realize the re calibration of the feature,so as to improve the feature representation ability of the model.3)The problem of low utilization of convolutional neural network features is further improved on the structure.A dual branch dense connection convolutional neural network model is constructed and combined with SVM.The dual branch structure makes the extracted feature information more abundant.The dense connection method integrates the shallow features with the deep features,and fully improves the deep convolution network's shallow features The reuse of layer features improves the accuracy of seed species identification.The research results of this paper can be used for crop seed variety identification and improve the recognition accuracy of existing seed sorting methods.And the simplified network structure is convenient for its deployment in the front-end equipment with limited computing resources,and has a wide application prospect in the sorting system.