Research On Fruit Detection Technology Based On Deep Learning

With the growing demand for fruits in China,the area of fruit production and planting has also been expanding.However,due to the decreasing number of people engaged in agricultural production in China and the low degree of automation of the Chinese fruit industry,this will have a certain impact on the Chinese fruit industry.In order to ensure the sustainable,healthy and rapid development of the fruit industry,intensifying efforts to develop related research on fruit picking automation technology is of vital help to the economic benefits brought by social development and the broad market prospects.Completing the detection of target fruits through computer vision technology is an important issue in the development of fruit picking automation.In recent years,with the tremendous breakthroughs in object detection with deep learning,More and more deep learning-based object detection algorithms have been proposed.Its powerful feature extraction ability makes deep learning-based object detection quickly become a hot research topic in various fields.In this paper,the regression-based object detection algorithm SSD(Single Shot Multi Box Detector)is used,and through the improvement and optimization of the classic SSD algoritm,the detection accuracy of small target fruits in natural environment is effectively improved.The main research work of this article can be divided into the following points:Firstly,make a suitable fruit data set,and increase the number of data sets by flipping,cropping,shifting and scaling in the data augmentation under the limited experimental image acquisition.After data augmentation,the sample of each type of fruit on the data set grows to 1500 pictures.Secondly,improve the classic SSD detection model to improve the accuracy of the model for small target detection.Replace the basic network VGG-16 in the classic SSD model with the residual network Res Net101,and build an improved SSD model according to the FPN,using multi-scale feature fusion to upsample the high-level features transfer to the low-level layer to obtain high-resolution,high-semantic feature maps,so as to improve the accuracy of the model to detect small targets.The improved SSD model was verified by experiments on the fruit data set,and the average accuracy of its target fruit detection reaches 84.34%,which is higher than the overall accuracy of the classic SSD model and reduces the detection of small object at the same time.Finally,optimize the improved model,and apply it to fruit detection in natural environment.In order to better fit the research of fruit detection in natural environment in this article,aiming at the occlusion problem of fruit in natural environment,in the improved SSD network model,the clustering algorithm is used to reset the aspect ratio of the default frame to reduce the training network parameters.Optimize the selection of candidate region for non-maximum suppression algorithm to improve the detection accuracy of fruits under occlusion.Experiment shows that the average accuracy of the improved SSD model after optimization reaches 88.15%,which is 3.81 percentage points higher than the model before optimization,and the detection speed of a single image is also reduced by 36.8ms.