Research On Intelligent Detection Algorithm Of Tea Buds Based On Deep Learning

Tea picking is a delicate work with high timeliness and seasonal requirements.The time and integrity of the buds are directly affected by the quality and price of the finished tea.Tea picking is currently mainly manual picking,supplemented by mechanical picking.Manual picking has the problems of high labor intensity,low efficiency,and high cost.Although machine picking can reduce economic costs and improve tea picking efficiency,it lacks the selection of old leaves and young shoots,and increases the workload of sorting tea stems.The automatic detection of tender buds is an important prerequisite for the realization of mechanical precision picking.The performance of the tea bud detection model is greatly affected by factors such as tea variety,light intensity,bud size and occlusion.Three kinds of tea bud images with different light intensities and different densities were collected as research objects.Through image brightness correction and network structure optimization,a deep learning detection model for tea buds that is easy to deploy in embedded systems was established,and a small program was developed to realize Automatic detection and result display of tea shoot images in natural environment.The main research process and results are as follows:(1)Tea image acquisition and preprocessing.During the spring tea picking period in 2020 and 2021,under different light intensity conditions,digital cameras were used to collect images of Longjing 43,Cuifeng and Anji white tea with different densities.Aiming at the overall whitening of high-brightness images and the inconspicuous contrast between young buds and old leaves,a regional brightness adaptive correction algorithm was proposed to improve image quality and enhance the significance of young buds features.After manual annotation of all samples,the number of training samples was expanded by three kinds of image processing: adding noise,enhancing contrast and horizontal flipping.(2)Establish a detection model of Longjing 43 tea buds based on deep learning.In order to verify the performance of different deep learning models for buds detection and to select detection models that are easy to deploy in embedded systems,four different models,Faster R-CNN,YOLOv3,YOLOv4 and YOLOv4-tiny were trained and tested under the same tea image samples and hardware environment,the accuracy of buds detection was above 90%.Considering that the intelligent tea picking machine has high requirements for model detection speed,model calculation complexity,and model parameter quantity,YOLOv4-tiny,which has a small model capacity,fast detection speed,and easy deployment,was selected as the basic network for buds detection.(3)Establish an improved YOLOv4-tiny tea buds detection model.In view of the problems of dense occlusion and missed detection of small-scale buds in the detection results of the YOLOv4-tiny model,the neck network of the YOLOv4-tiny network model was improved.By introducing a large-scale feature layer of 52×52,the attention degree of the feature information of small target buds was improved,and the attention mechanism was added to suppress background noise,and the salience of small-scale buds features was improved,and the two-way feature pyramid Bi FPN was introduced to fully integrate different scales feature information.Ultimately,the detection performance of the model for dense occlusions and small-scale targets was improved.An ablation test was performed on the improved mechanism to verify the effectiveness of the improved strategy.The detection precision rate of the improved YOLOv4-tiny-Tea model was 94.7%,and the detection recall rate was 87.3%.Compared with the YOLOv4-tiny detection results,the detection precision rate and recall rate have increased by 4.4 and 13.7 percentage points,respectively.(4)Performance Analysis and Results Visualization of a Tea Sprout Detection Model.In order to verify the detection performance of the model,different light intensities and two other tea images are tested.Compared with the original high-brightness shoot image detection results,the recall rate and F1 value of the high-brightness image shoot detection after regional brightness adaptive correction were increased by 14.6 and 7.8 percentage points,respectively.The detection accuracy of the YOLOv4-tiny-Tea model under different light intensities was 94.8% and the recall rate was 91.7%.The detection F1 values on Cuifeng and Anji white tea were 88.0% and84.0%,respectively,indicating that the established bud detection model has good robustness and generalization ability.Develop a We Chat applet to visualize the results of intelligent detection of tea buds.