Research On Fruit Image Target Detection Based On Convolutional Neural Network

Since the reform and opening up,China's fruit planting area has been expanding and fruit production has been increasing.However,in the past,most fruit picking operations relied on manual labor,while the number of people engaged in agricultural production has been shrinking.In order to cope with the inexorable trend of labor shortage in agricultural production,the development of picking robots with recognition and positioning functions meets the needs of social development.Social development has brought enormous economic benefits and broad market prospects.In recent years,with the rapid development and rise of in-depth learning,especially in the visual direction,it has made tremendous achievements.Compared with traditional recognition algorithms,in-depth learning is more powerful in describing the features of target recognition.Therefore,it is of great significance to develop a picking robot based on the visual recognition function of deep learning algorithm.Firstly,based on the current advanced deep learning algorithm,this paper proposes a fruit image classification and recognition algorithm based on convolution neural network.Referring to the classical convolution neural network model LeNet-5 structure,this paper proposes a new convolution neural network structure and classifies and identifies five kinds of fruits: apple,pear,orange,orange and peach.The model constructs an input.Layer,two convolution layers,two pooling layers,two full connection layers and one output layer.The experimental resultsshow that the proposed convolutional neural network structure not only achieves 96.88% recognition accuracy on the data set,but also has higher accuracy and faster convergence speed than the original LeNet-5 model.Secondly,the fruit image recognition is further studied in this paper.In order to demonstrate the feasibility of the algorithm,apple image is selected as the specific research object based on the deep learning target detection algorithm.At present,the mainstream deep learning target detection algorithms are Faster-RCNN,YOLO,SSD.The detection algorithm in this paper adopts the Faster-RCNN based on the region suggestion,through which the algorithm can be used.The recognition and location of apples in images are studied.In order to cope with the possible occlusion,overlap,backlighting and uneven illumination of Apple targets in natural scenes,Apple images with different sizes,quantities and different illumination angles were captured during image acquisition.The improved LeNet-5 and classical convolution neural networks VGG16 and ResNet101 proposed in this paper were selected to extract and model the network.By setting different hyper-parameter combinations and comparing the accuracy of the models,the appropriate hyper-parameter combinations are obtained.Finally,the average detection accuracy of 90.91% is obtained on the Faster-RCNN Apple detection model based on ResNet101,while the detection accuracy obtained on the improved LeNet-5 model proposed in this paper is too low and phase-sensitive.Compared with the model based on VGG16,the model based on ResNet101 improves the average detection accuracy obviously.Although the detection speed is lower than VGG16,and the time of detecting an image is 0.39 seconds,it has met the real-time requirements,and obtained 98.96% recall rate and 85.74% recall rate.The final experimental results show that the model is available.Thefeasibility of the algorithm is verified by the detection of apples on trees.It also provides experience for the research of other kinds of fruits.