Research On Fruit Image Recognition Algorithm Based On Deep Convolutional Neural Network

China is an important fruit producing country in the world.The complexity of the fruit growing environment makes the current fruit picking and processing operations still rely on labor.However,with the development of urbanization,rising labor costs and a sharp decline in agricultural employment make the fruit industry face the reality of labor shortage.Fruit picking robots can improve production efficiency and solve current problems.Fruit image recognition is an important part of the fruit picking robot vision system.In recent years,deep convolutional neural networks have developed rapidly.Compared with traditional image recognition algorithms,it have stronger ability to express target features and better recognition results.Using convolutional neural networks to identify fruit images can provide strong technical support for the development of fruit picking robots.This thesis mainly studies the algorithms of fruits image detection and classification based on deep convolutional neural networks.Firstly,in order to solve the problems of low accuracy and fewer types of fruit recognition methods,meet image recognition requirements,a fruit image classification method based on improved Darknet-53 was proposed.In this method,after analyzing various normalization methods,the group normalization is used to replace the batch normalization in the original network structure to construct a Darknet-53-GN network.At the same time,22 kinds of fruit classification data sets were constructed.The results show that the method is not affected by the batchsize and achieves 95.6%classification accuracy.Secondly,a neural network model for fruit object detection is designed.Based on the YOLOv3,the GIoU bounding box regression loss function is adopted,and the idea of residual link skip connection is introduced in the YOLO layer of the original network for connection.The trained Darknet-53-GN classification network is used for weight initialization the fruit object detection network model,then train the network model with self-built labeled fruit datasets;the average accuracy is 90.85%,and it has good robustness for fruit object recognition under the influence of different natural environment.Finally,a lightweight fruit object detection model is designed for the portable mobile device application scenario.MobileNetV3-Large,a lightweight network proposed based on the mobile terminal,can effectively reduce the complexity of the model,and is thus used to replace the backbone network Darknet-53 in the original network model.And use the GIoU bounding box regression loss function.The memory occupied by the proposed model is 32 MB,the size of the model is reduced to one-seventh of the original.The inferred speed of a single picture on is 36.85 ms.The accuracy of the method is almost constant,which significantly improves the detection speed and increases the availability of the model.