Detection And Classification Of Damaged Wheat Based On Neural Architecture Search

Grain is one of the main sources of energy for human bodies.Quantity and quality of grain are both closely related to national development and social stability.Grain is often damaged due to mildew,pests and other factors during the storage process,which will cause the quality of stored grain to decrease.Grain damage will not only lead to an increase in the global hunger rate,but long-term consumption of grain infested by molds and pests will lead to human malnutrition and even induce diseases.Timely detection and separation of damaged grain can not only reduce the loss of stored grain,but also avoid human diseases.Therefore,the grain quality inspection work is very important and urgent.In this study,mildew-damaged wheat kernels（MDK）,insect-damaged wheat kernels（IDK）and Undamaged wheat kernels（UDK）were studied.In this paper,we propose two methods based on neural framework search for detection and classification of damaged wheat.Both of the neural networks generated by the two methods have achieved good results in the detection and classification of damaged wheat.The experimental results prove that the two methods proposed in this paper are feasible and effective in the detection and classification of damaged wheat.At the same time,this paper is the first time to apply the neural network generated by neural architecture search to the field of grain detection,which brings new methods and ideas for other fields.The research contents of this paper are as follows:（1）Firstly,the collected impact acoustic signals of wheat kernels were transformed into spectrograms.Then,the problem of too few data sets is solved by five methods:signal endpoint detection,adding noise,Gaussian blur,adjusting contrast and brightness,and deep convolution generation adversarial network.Finally,the composition of the data set in this paper is summarized.（2）For the MDK,the IDK and the UDK,a neural framework search method was proposed for the first time to detect and classify damaged wheat kernels.Progressive neural architecture search is to traversal cells in the search space based on the sequential model-based optimization,and then predict the cell structure with the best performance through the Predictor function.Finally,a complete neural network is constructed by stacking cells with preset methods.The experimental results show that the average F1 value of the three types of wheat kernels is 96.2%,which is higher than the popular artificial neural network.It is verified that neural architecture search can be used to generate neural network,and the performance of the obtained neural network is excellent.（3）In order to solve the problem of large memory cost and long search time of progressive neural architecture search,a progressive partial channel connected differentiable neural architecture search method is proposed,which generates more stable neural network model with less time.Differentiable neural architecture search is to continuously relax the discrete search space to obtain a differentiable learning target,and then use gradient descent method to generate cell structure.In this paper,the progressive partial channel connected differentiable neural architecture search is proposed,which reduces the memory consumption by sampling a small part of channels on the basis of differentiability,and gradually increases the depth of the neural network to reduce the performance gap between search and evaluation.Finally,the neural network generated by the proposed method was applied to data sets of three kinds of spectrograms of wheat kernels,and excellent classification results with average F1 value of 96.6%were obtained.This result not only surpasses the performance of the artificial neural network,but also surpasses the performance of the network generated by the progressive neural architecture search.