Research On Image Recognition System Of Wheat Grain Integrity Based On Deep Learning

Wheat is the third largest food crop in China,and it occupies an important position in our country's food consumption.70% of the wheat produced in our country each year is used to make wheat flour,which is then processed into traditional foods such as noodles,steamed bread and dumplings.Therefore,the quality of wheat is closely related to food safety and consumer interests.In the process of wheat trading and circulation,the detection and grading of wheat quality is an indispensable part,which is the basis for subsequent wheat pricing,and has important practical and strategic significance.At present,the detection of wheat grains appearance is mainly done by quality inspectors through human vision,which lacks objectivity and scientificity.Most of the domestic commercialized wheat detection facilities use traditional machine learning algorithms as the recognition algorithm,and their accuracy rate needs to be improved.These facilities have some problems such as strict detection environment and low detection efficiency.In this paper,deep learning technology is used to identify the integrity of single wheat grain,detect objects of multiple wheat grains,and develop an image recognition system of wheat grain integrity.The main research contents and conclusions are as follows:(1)The integrity of single wheat grain was identified based on deep learning and image processing technology.Images of sound and broken grains of 9 kinds of wheat were collected,and were preprocessed through gray processing,binarization processing,morphological processing,contour extraction,image segmentation,image denoising,image enhancement and data expansion.An image data set of single wheat grain was built.With the help of label Img which is an image annotation software,the integrity and position of all wheat grains in the image were labeled,and then the image data set of multiple wheat grains was built.The training set,validation set,and test set 1 of the two image data sets are composed of 6 kinds of wheat,and the test set 2 is composed of 3kinds of wheat.Four kinds of convolutional neural networks were improved respectively to construct W-Alex Net,W-VGG-16,W-Goog Le Net and W-Res Net-50 network models for identifying the integrity of single wheat grain.The four network models were trained on the Google Colaboratory platform using Adam optimizer with learning rates of 0.0005 and 0.0001.The test results showed that image preprocessing and transfer learning effectively improved the convergence speed of the model,and increased the recognition accuracy by 1% and 2.5%,respectively.Among the four network models,W-Goog Le Net had the lowest recognition accuracy,about 78%.W-Alex Net had the fastest recognition speed,and it took about 0.02 s to recognize a single image.W-Res Net-50 had the best recognition performance and its recognition accuracy of test set 1 and test set 2 were 98.50% and 99.17%,respectively.Its recognition time of a single image was about 0.13 s,and the size of its network weight file was moderately 94.4MB.(2)The objects of multiple wheat grains images were detected based on deep learning.The traditional SSD object detection network was improved to construct W-SSD network model based on Res Net-50.After training with the Adam optimizer with a learning rate of 0.0001,the model's m AP on the validation set reached 1.0,and the m AP curve was relatively smooth.The accuracy of the model on the test set reached99.11%,the missed detection rate was only 0.11%,and the detection time for a single image was about 0.34 s.The detection result images show that the W-SSD network model accurately marked the integrity categories and location of wheat grains,indicating that the model had reliable recognition capability,good generalization and applicability,and could be applied to object detection of multiple wheat grains.(3)The image recognition system of wheat grain integrity was developed.In order to realize the batch and continuous real-time detection of wheat grains,the hardware of the image recognition system composed of feeding device,conveyor,USB camera and computer was designed and constructed,and an interactive image recognition software for wheat grain integrity was developed.The test results showed that the image recognition system had an average counting error of 1.6% for 50 sound grains and 50 broken grains,an average recognition accuracy rate of 99.05%,and an average detection time of 70.8s.The recognition results are visualized and output to reduce the workload of quality inspectors.In the process of identifying 50 g wheat grain samples,the recognition efficiency of the image recognition system was higher than that of manual detection,with an average time of 601.8s and an accuracy rate of 98.97%.The system has good recognition performance,meets the requirement of actual detection,and provides technical support for wheat detection facility based on deep learning.