| Material information collection is an essential part of the automatic logistics system.Information collection requires fast and accurate.At the same time,the existing information on the logistics package should be used for collection to the full in order to reduce the intermediate links and human intervention.At present,manual input,barcode or RFID are widely used in logistics system to collect material information.However,with limited types of goods in food,medicine and other industries and no information carrier such as barcode or radio frequency card posted on the external surface,for this reason,it is necessary to develop an intelligent sorting system with high speed and recognition rate to directly identify the cargo category information.Based on the digital image processing and convolutional neural network,in this thesis,Open CV and Tensor Flow are used to realize,Qt is used to development interfece,and an intelligent sorting system is developed independently.The whole system is constructed with modular design idea,which has strong applicability.The thesis is organized as follows:Firstly,according to the automatic sorting system and the general process of goods sorting,the image acquisition platform was designed to obtain goods images,and the selection basis of the photoelectric sensor,camera,lens and light source that constitute the image acquisition platform was in introduced.Secondly,the method of segmenting goods image on the logistics conveyor line is studied.First,in order to obtain a suitable binary image,the original color image is processed by grayscale,median filtering,Gamma correction,Otsu method binarization and closed operation to reduce the adverse effects caused by noise,uneven illumination and reflection of the belt thermal sintering connection part.Then,the outer contour of the goods is found by process and analyze the binary image,and the target area.is extracted from the original image on the basis of bounding box.The result indicates that the whole object area can be segmented in the logistics industry environment.Thirdly,the images of 13 kinds of goods are collected on the image acquisition platform,and the data set is expanded by rotation and trained with Alex Net,VGGNet,and lightweight CNN.The result shows that the network performance can be optimized by adjusting theparameters and deepening the network structure,however,the network layer is not as deep as possible.In fact,the verification set turns out completely correct when using lightweight CNN to train 360 degree rotation image in this thesis.Forth,according to the mechanism research of CNN,feature extracted from low-level to high-level in CNN,and CNN has poor robustness to rotated images.The result has been confirmed experimentally in this thesis.If only the pictures of objects placed in fixed direction are collected for training,and the pictures of objects rotating 45 degree direction are used for testing,it is found that some categories are recognized correctly,and some categories are recognized incorrectly after rotating 45 degree.The experimental proves that CNN has poor invariance to rotation,and the self-build goods data set is expanded by rotation can avoid overfitting and ensure the accuracy of goods image recognition at any angle.Finally,Image processing and CNN training process are realized with Open CV and Tensor Flow,and system is developed on portable platform Qt.After test,the time from camera to system identification and display of goods category is 400?600ms with relatively high accuracy,which can completely meet the needs of logistics automation.The whole intelligent sorting system is developed by module.For different projects,the code of image segmentation part and building convolution network part be modified is unnecessary orexclusive,so it can be used quickly.Moreover,the automatic identification of many kinds of goods on the automatic logistics transport line can be realized by adding different types of goods to the training with extensibility in the model. |
PDF Full Text Request