Research On Key Techniques Of Apple Sorting Based On Machine Vision

With the rapid development of industrial automation,robot sorting technology has been applied to various fields of production and processing.At present,in the field of fruit sorting,the production line used in the actual production is usually only for single index,single type of fruit sorting,the sorting effect is not ideal,and the machine idle rate is high.Robot sorting can easily achieve multi index and multi type fruit sorting,and easy to implement product traceability.Two key technologies of identification and location are studied in this paper.First,according to the functional requirements of the sorting system,the overall plan of apple sorting system is designed.The system is mainly composed of visual inspection module,conveyor belt module,robot module and communication module.The hardware selection of the sorting system is completed,mainly including the light source,the camera and the robot system and so on.The image collection box is designed and built,and a feasible system software design scheme is proposed for the apple sorting system.Secondly,in order to realize the recognition and classification of apple,this paper takes red Fuji and red Marshal apple as the research object.First,we propose a multicore least squares support vector machine(SVM)which uses linear kernel and RBF kernel to automatically search kernel weight.Two kinds of Apple classification models and the classification model of red Fuji apple are set up respectively.Then the deep convolution neural network algorithm is studied,and a depth learning algorithm based on LeNet-5 improved neural network model is proposed.The algorithm adds Flatten layer to color image,flattened the three channel data into single channel,and then adds the Dropout strategy to the full connection layer to increase the generalization ability of the model.The LeakyReLU function is used instead of the original Sigmoid activation function to alleviate the gradient dispersion and gradient disappearance caused by the increase of the number of layers.The results of the comparison of two methods in classification and classification of apples show that the SVM method has a better classification effect compared with the CNN method when the number of Apple images is less,and when the image data is sufficient,the CNN method is obviously better than the SVM method in both the training and test time or the classification accuracy.Finally,the location method of apple is studied,and an improved Hough circle detection algorithm for determining the central coordinates of the apple is given.And combined with the improved Zhang Zheng you calibration algorithm,the robot's grasp and location of Apple target is realized.The results show that the improved Hough transform circle detection method is more accurate than the apple center of the original Hough detection,but the operation efficiency is not high,and it is not suitable for the high resolution image processing.