Research On Robot Recognition System Based On Deep Learning And Information Fusion

For the robot fruit sorting line,there are problems such as single sorting object and low accuracy,the subject used information fusion and deep learning recognition methods to study it.Firstly,information fusion method had been used to extract the color,shape and gravity information of the object through image sensor and gravity sensor,and the information was fused at the feature layer to obtain the fused feature vector,and BP neural network was established to classify and identify it.Then the deep learning method had been used to study,had extracted the deeper feature information of the image,had built a convolutional neural network model,and fruit had beed classified.This method is more accurate than the information fusion method,and the correct rate can reach 97.33%.In the subject,the convolutional neural network used image sensors,which improved the accuracy of fruit sorting and saved the cost of use.The method is portable and can be transplanted to other sorting fields.Based on BP neural network,the information fusion fruit sorting system was built.The color and shape features of the image was extracted,and the gravity characteristic information of the object was collected by the gravity sensor.The information was fused in the feature layer,and the fruit classification model was built by the neural network.The fusion feature was trained and learned to realize fruit classification.The deep confidence network,cyclic neural network and convolutional neural network had studied in deep learning.Based on Python development environment,Tensorflow image recognition framework and Tkinter visualization tools,a convolutional neural network image recognition system had been built.It included user management module,image acquisition module,data set making module,model building module,model training module,model optimization module and image recognition module.For the multi-object recognition situation,the improved image segmentation algorithm had been used to quickly segment and locate the image to meet the real-time segmentation and positioning requirements of the image.The camera has been calibrated,the camera internal parameters had been obtained,the camera external parameters had been set according to the actual installation position of the camera,and the pixel coordinates had been converted into the world coordinates of the object by coordinate transformation to realize object positioning.The robot fruit sorting experiment platform was built,and the segmented fruit image was identified by convolutional neural network image recognition model,and the identified fruit was positioned and captured.The experimental results show that the segmentation algorithm is fast and accurate,and the recognition model has the accuracy and accuracy of object location,which proves the feasibility of deep learning robot fruit sorting.