Research On Rice Seed Image Classification Algorithm Based On Deep Learning

As a crucial branch of image processing,image classification has high academic value.With the rapid development of deep learning,image classification research based on deep learning is being applied to various scenes of people's lives.In agriculture,hybridization technology has made crop varieties more and more diverse.Effective and accurate judgment of crop seeds has become a challenge.Therefore,it is of significance to introduce image classification algorithm based on deep learning in crop seed classification field.Rice is one of the vital food crops in China.With the development of agricultural technology,rice varieties are becoming more and more abundant.The mixing of low-quality seeds and high-quality seeds is inevitable.Hyperspectral imaging technology combines imaging technology with spectral technology to effectively obtain multi-dimensional information of samples.Compared with ordinary images,hyperspectral images have higher spectral resolution,so they are more suitable for fine classification of samples.This paper combines hyperspectral imaging technology and deep learning-based image classification algorithms to jointly solve the rice seed image classification problem.In deep learning,the ability of a model to learn features will directly affect the final result of image classification.For rice seed samples whose class features are not significant,targeted learning of effective features can improve classification accuracy.In this paper,SENet structure is introduced to fully extract the sensitive features with large differences between classes for learning,and suppress other features that are not important for rice seed image classification.Using SE-ResNet,AlexNet,and VGG three network structures for model selection comparison experiments,the experimental results show that the classification results of the three network models are quite different.The rice seed samples used in the experiment have small differences between classes,and it is difficult to distinguish them with a simple network structure.Therefore,SE-ResNet with the best performance is selected as the model basis for the improved algorithm.Based on the selection of the SE-ResNet model with superior performance,this paper analyzes the working principle of the loss function in detail and proposes an improved joint loss function based on Cross Entropy Loss.It is a commonly used loss function in multi-classification tasks.According to its characteristics and structure,this paper defines a kind of auxiliary loss function which is similar to its structure but different in function,and introduces a variable parameter to combine the two.In the training process,Cross Entropy Loss as the main loss function,always controls the trend of the model,and the auxiliary loss function regulates the network within a certain range,which reduces the difference between classes,increases the loss value,increases the learning difficulty of the network,forces the network to learn the features of the greater difference between classes and the smaller difference within classes,so as to improve the performance of the model and improve the classification accuracy.The experimental results prove that the proposed improved joint loss function is beneficial to make the network model fully learn the subtle features between rice seed images.Finally,the test accuracy of classification on 6 types of rice seeds with high similarity is 94.77%.