| Pecans are rich in nutritional value and are important economic crops in the mountainous areas of southern Anhui.However,the internal structure of pecans is complex.It is difficult to maintain the integrity of the kernel during the process of breaking the shell and taking the kernel.In order to more effectively maintain the integrity of the kernel and reduce the damage of the kernel,it is necessary to classify and process the hickory materials after the shell is broken once,so as to improve the level of deep processing.The pecans after breaking the shell are dusted and screened,the shelled materials are subdivided into 5 categories:relatively complete shell kernels without separation,dew kernels,unbroken pecans,shells,incomplete shells without separation.For research convenience,set A,B,C,D,E as its labels in the paper.In the research,the method of deep learning combined with computer vision was introduced to classify the hickory mixture after the shell was broken.The main work of this paper is summarized as follows:Making image acquisition platform,a total of 1713 samples of 5 types of sample images were collected,and the images were preprocessed and data enhanced,and 15,000 images were generated to make a Dataset.The classification model is constructed based on the VGG16 network,and the data set is trained and tested.The results show that the convolutional neural network can achieve 99.5% accuracy in the classification of broken shell materials.Detect images containing multiple materials,Use Label Img software to annotate images,Make a target detection dataset,build SSD and YOLOv4 two regression-based target detection models,train and test the models,and test the material detection of the models the results show that the m AP values of the models are 96% and 97.1% respectively.The average detection accuracy of the two models is not much different,but for the detection effect of incomplete shell kernels and undivided materials,YOLOv4 is better,so YOLOv4 is determined as the online recognition score.The selected model.(3)According to the online identification and sorting process,determine the material separation method,design the overall structure of the system,and select the main components.Load the trained model to the online recognition and sorting platform,and carry out static recognition test and dynamic sorting test respectively.The static recognition rate is 95.6% and the dynamic sorting rate is91.4%,achieving a good sorting effect.According to the online identification and sorting process,determine the material separation method,design the overall structure of the system,including conveyor belt,host,electrical cabinet,camera,solenoid valve,etc.and select the main components.Load the trained model to the online recognition and sorting platform,and carry out static recognition test and dynamic sorting test respectively.The static recognition rate is 95.6% and the dynamic sorting rate is 91.4%,achieving a good sorting effect. |
