Research On Tea Tree Bud And Leaf Recognition Based On Generative Adversarial Network

The recognition and detection of tea tree buds and leaves is necessary to realize intelligent picking of tea tree buds and leaves,and the recognition of tea tree buds and leaves using deep learning methods has become a new trend.Since tea tree buds only appear in spring,the collection time is very short,and the tea tree buds in tea gardens are widely distributed,the difficulty and cost of tea tree bud image collection is high,making the tea tree bud image dataset relatively small,which limits the research and application of deep learning on it.Therefore,it is necessary to expand the tea tree buds dataset,and the traditional data enhancement methods have certain shortcomings and no diversity in the augmented data samples.The emergence of generative adversarial networks can provide a new solution to the data shortage,but the original generative adversarial networks have problems such as unstable training,insufficient feature extraction and insufficient quality of generated images when generating data.Therefore,this thesis improves the deep convolutional generative adversarial network(DCGAN)to generate high-quality tea tree buds and leaves images,expands the tea tree buds and leaves dataset,and validates it with the mainstream target detection model,improves YOLOv5 to improve the detection accuracy and reduces the model size.The research results show that the image data expanded by generative adversarial network can not only make up for the shortage of tea tree buds image dataset,but also improve the accuracy of the target detection model,which provides a certain basis for constructing tea tree buds image dataset and realizing intelligent tea tree buds picking.The main research contents of this thesis are as follows:1.To address the problem of insufficient data of tea tree buds and leaves images,DCGAN is used as the base network,the residual network structure is used in DCGAN,the activation function and loss function are improved and the attention mechanism is introduced.To evaluate the performance of GAN,Inception Score(IS)and Fréchet Inception Distance(FID)are used as objective indicators to evaluate the quality of generated samples.The results of the ablation experiments demonstrate that the improved model is better than DCGAN in both FID and IS.the original DCGAN model has a FID of87.37 and IS of 2.375,while the improved DCGAN-Im model has a FID of 153.81 and IS of 3.213.2.In order to improve the training speed of the model and reduce the computation of the whole model,the tea tree bud detection model was constructed based on YOLOv5,and the Ghost-Bottleneck module replaced the original Bottleneck module in YOLOv5.The accuracy of the improved YOLOv5-Im is experimentally verified to be 3.49% higher compared to the original model,and the model size of the improved YOLOv5-Im is 28%lower than that of the original YOLOv5,and the floating point operation is 33% lower than that of the original YOLOv5.3.Validate the effectiveness of the data enhancement algorithm.The experimental results show that the accuracy of the target detection model trained by the hybrid data set is improved by 3-4 percentage points.This demonstrates the effectiveness of the data augmentation algorithm to enhance the performance of target detection models by expanding the samples.