Research On Detection Of Lingwu Long Jujube Based On Deep Learning

As an important economic fruit in Ningxia,Lingwu long jujube has been vigorously developed locally.So the planting area has gradually increased every year.In order to reduce the labor of traditional inefficient manual picking and to achieve efficient automatic picking technology,this paper focused on a variety of algorithms based on deep learning target detection are used to study the detection and location technology of Lingwu long jujube in automatic picking.The research contents and results are as follows:1.Collection of an image data set of Ningxia Lingwu long jujube(2000 pieces),including a variety of images taken from different angles such as one jujube,two jujubes,multiple jujubes,leaf occlusion,jujube adhesion,long jujube smooth light shooting and inverse light shooting.According to the experimental requirements,the image set is preprocessed and labeled uniformly,and the data set containing the source image path,size,depth and coordinates of the location of the long date is obtained.2.The target detection model based on deep learning candidate region algorithm was used to detect Lingwu long jujube.Based on the deep convolution neural network CNN,the Lingwu jujube detection network of RCNN,Fast RCNN and Faster RCNN models was designed,and each long jujube detection model was obtained by network training.The experimental results show that the above three algorithms can detect the coordinates of the dates in the original image.When the image contains a single or two jujubes,the detection effect is best,and the detection rate is an average of 80%.When the image contains multiple long dates that are blocked and glued,the detection rate will be significantly reduced.3.The SSD target detection model based on the deep learning regression algorithm was used to detect Lingwu long jujube.The SSD algorithm has regression idea,which simplifies the computational complexity of the network and improves the detection speed and accuracy.Based on SSD,the detection model of long jujube is obtained by the original SSD network training.The test experiment shows that this model realizes the detection of long jujube with an accuracy of 90%.The detection results were found to be less accurate or not detected when the image was dark,and the detection rate was also affected when the image contained occlusion and adhesion.4.In response to the above problems,this paper proposes a long date detection model based on improving the SSD network.The improvement point is first to enhance the brightness of the training set image,and then to standardize the basic network for batch processing,and to perform multiple convolution processing deep in the network.Finally,the model parameter match threshold and dropout keep probability are optimized.After the improved training,a new SSD long date detection model was obtained.The detection rate of this model has been significantly improved,and the detection accuracy is as high as 93%.When the image contains occlusion,adhesion,etc.,the detection effect is better than that of the original SSD model.Comparing the above five models,we come to the conclusion that these five models have achieved the task of detecting Lingwu long jujube.Among them,the long dates detection model based on improving the SSD network has the best effect,and its detection accuracy is as high as 93%.It can be proved that it is feasible to apply deep learning target detection algorithm to the detection of Lingwu long jujube,which is a good basis for realizing the precision detection and picking of Lingwu long jujujube by automatic picking robot.