| Modern agriculture requires the automation and intelligence of agricultural machinery,and the emergence of picking robots lays the foundation for this.The main research content of this subject is the study of mango recognition binocular stereo vision system based on convolutional neural network.Through image preprocessing after binocular vision image acquisition,the deep learning of convolutional neural network intelligent algorithm is used for feature extraction so as to identifying mango target effectively.When it comes to identifying mango targets in complex environments,due to the limitations of environmental factors and mango’s own characteristics,the extraction ability of traditional methods is insufficient.To this end,feature information is further extracted and enriched by adding edge prediction of image.Then,in order to achieve the deep processing of feature information,the quantification processing of image feature information is realized through image feature coding and convergence with the image enhancement technology so that the feature information of the target is saved to the greatest extent,which lays a foundation for subsequent target recognition.In machine learning,convolutional neural network is a deep feedforward artificial neural network,which has been successfully used in image recognition.In the algorithm design of the convolutional neural network,due to the diversity and difference of of structures,many algorithm models are derived,which have their own characteristics as well as advantages and disadvantages.Therefore,finding an efficient identification algorithm suitable for mango target is the key to solve the problem.Therefore,in various convolutional neural network algorithms,R-CNN(region-convolution neural network),DenseNet(Dense network)and YOLO(You only look once)are compared and analyzed in detail,and We finally decided to improve the YOLO algorithm,and take it as the core algorithm of this experiment.In order to make it more suitable for mango target identification,three aspects of applicability improvement for the current identification problem have been proposed in this thesis.There are mainly improvements in network depth,introduction of sampling layer based on spatial pyramid and improvement of network training strategy.Finally,the algorithm complexity is analyzed and compared with other network models to verify the feasibility and superiority of our improved algorithm.In the design of the test plan,in order to ensure the applicability of the system,we ensure the diversity and integrity of the sample information by adjusting the environmental factors such as the time and place of image acquisition during the collection of target sample information.Finally,the experimental data processing of the mango recognition system is tested.The identification data of the improved YOLO algorithm,the original YOLO and R-CNN algorithms are compared and analyzed.By comparing the accuracy and recall rate of each algorithm,the improvement of recognition accuracy of the improved YOLO algorithm is verified,and the advantage of recognition is demonstrated when dealing with covered or overlapped mango images. |
