Research On Rice Seeds Defect Recognition And Thousand Kernel Weight Measurement Of Overlapping Distribution Based On Deep Learning

With the continuous growth of the global population,it is also necessary to maintain the availability and high quality of grain on the premise of ensuring a substantial increase in grain output.The detection of seed quality is of great significance to the yield increase and high yield of crops.Rice is the main crop in Asian countries,and more than 50%of the world's population takes rice as its staple food,which has great economic value.The identification of rice seed defects is the basis of thousand kernel weight measurement.The sample data of the two is an important index to evaluate the quality of crops and check whether the breeding work has achieved results.Existing image processing methods have some shortcomings,such as fixed position of collected pictures,complex processing process,inability to normalize processing,etc.Therefore,depth learning classification and target detection methods with strong generalization and adaptability can solve the above problems.The main content of this paper includes the identification of multiple defects of rice seeds and the measurement of thousand kernel weight under overlapping distribution.Combined with the characteristics of Wuyunjing-24 rice seeds,targeted pretreatment and the establishment of a correlation model provide technical guarantee for the identification of defects of rice seeds and the measurement of thousand kernel weight under overlapping distribution.In order to meet the needs of the establishment of deep learning database,this paper designs an intelligent rice seed data acquisition system.The system uses stepper motor and seed metering device to control the output of rice seeds,TXG12c industrial camera to obtain the image of defects and overlapping distribution,YZC-1B weighing sensor and LDST-I(V)-SP precision weight transmitter to obtain the sample weight information.The system has the characteristics of real-time detection and simple operation,which provides basic data for the research of rice seed defect recognition and overlapping distribution thousand kernel weight measurement technology based on deep learning.Similar features of rice seeds will increase the difficulty of recognition.Therefore,this paper proposes enhanced individual characteristics normalized lightweight Rice-VGG16 method for rice seed defect recognition.Firstly,rice seed defects are divided,and the image processing steps are used to standardize the seed images and construct the datasets.Secondly,the fifth max-pooling layer is modified to the ave-pooling layer,and the activation function is defined as Leaky Rectified Linear Units(Leaky-ReLU)to enhance the individual characteristics and improve the recognition accuracy.Then,a batch normalization layer is added after the last convolution layer of each convolution group,the first full connection layer is removed,the node number of the second full connection layer is modified to 1024,and the model parameters are fine-tuned to carry out model lightweight.Thus the normalized lightweight Rice-VGG16 model is constructed to improve recognition speed.Experimental results with real datasets demonstrated that:the model was able to accurately identify rice seed defects,with the training accuracy of 99.63%and the recognition accuracy of 99.51%.The existing image processing methods can only deal with the problem of single-layer seed counting well,and the overlapping state is often misjudged,thus reducing the accuracy of thousand kernel weight detection.To solve the above problems,this paper proposes a deep learning optimization method based on grouping and pre labeling of rice seed contour.deep learning optimization method based on contour grouping pre-labeling for counting overlapping rice seeds.First,the contour grouping method-based on the Euclidean distance and divergence function as comprehensive criteria-pre-labels the rice seed contours to reduce the number of potential false classification edge points.The purpose of integrating the results of contour grouping and pre-labeling in its feature extraction layer and using a linear combination of inter-class and intra-class error functions as the total error function.It can not only provide richer features at the Faster R-CNN input,but also reduce the labeling workload.Second,based on the improved Faster R-CNN deep learning optimization method(TKW-VGG16),the overlapping rice seeds were accurately counted.An orthogonal test is proposed to optimize the parameter configuration of the full connection layer,aimed at improving the speed of deep learning.The experimental results demonstrate that the average error rate of rice seeds in a single image is 1.06%and the average recognition time of counting was 0.45s.With the improvement of computer processing capacity and the continuous development of measurement methods and methods,the application of rice seed defect identification and overlapping distribution thousand kernel weight measurement technology based on depth learning can greatly improve the efficiency of seed examination,and at the same time provide technical support for seed examination of grain seeds with small targets and multiple similar characteristics.