Research And Application Of Rice Chalkiness Recognition Algorithm Based On Deep Learning

Chalkiness is an important factor affecting the processing and food quality of hybrid rice.Accurate statistics of the chalky region information are of great significance to the relevant agricultural departments,and the recognition of chalkiness is a key step.The existing recognition algorithms of chalkiness are mainly based on traditional image processing algorithms,while the actual rice images have more touching seeds and the color and texture information of the chalkiness is very complicated.The algorithms based on artificial rules are used in the recognition of chalkiness,and the performance is not ideal.In this thesis,a rice chalkiness recognition algorithm based on deep learning is studied to improve the accuracy of chalkiness recognition.The main research contents are as follows:1.This thesis studies a segmentation algorithm of touching seeds based on concave points matching.Since the neural network designed in this thesis is aimed at single seed,the touching seeds should be segmented firstly.Combining morphological operations,concave points matching and distance transform function,this thesis improves the existing segmentation algorithm from three aspects.From the aspect of simplifying the complexity of the problem,the error rate of the concave points matching is reduced by reducing the number of concave points.From the aspect of optimizing the matching rules,the matching search range of the concave point pair is optimized,and the efficiency of the algorithm is improved while ensuring the segmentation accuracy.From the aspect of overcome the shortcomings of the algorithm,a segmentation algorithm based on the distance transform function is proposed to solve the isolated concave points that the algorithm could not deal with.handle ultimately improve the segmentation effect of the algorithm.The algorithm achieves effective segmentation for complicated scenarios of the touching seeds.2.This thesis designs a convolutional neural network G-Chalk dedicated to the recognition of single rice chalkiness.According to the characteristics of low resolution of rice image and large differences in size of different parts of rice,this thesis improved the FCN-8s semantic segmentation network from three aspects: fusion of feature map,deconvolution and loss function,and design a convolutional neural network G-Chalk which is suitable for rice chalkiness recognition.Due to the lack of rice database forsemantic segmentation,a single grain rice database Chalk4 is constructed to train the network model and verify its effectiveness.In the end,the optimal parameters of the G-Chalk network are selected by combining a large number of contrast experiments,which greatly improve the network's recognition accuracy compared with the existing algorithms,and algo has stronger adaptability and robustness.3.In this thesis,a SLIC superpixel segmentation algorithm based on GLCM is studied.To solve the shortcomings of SLIC superpixel algorithm in dealing with low-resolution image,this thesis improves the SLIC algorithm based on the image texture information obtained by GLCM,so that it can get the super-pixel with more accurate edge information.The fine edges obtained by the SLIC algorithm overcome the shortcoming of rough edge extraction of semantic segmentation network,which further improves the recognition accuracy of the overall algorithm.4.This thesis design and implement a rice quality analysis software for statistical information of rice.The software provides a good human-computer interaction experience,enabling users to efficiently recognise chalky areas in rice images.