Research On Few-Shot Learning Methods For Fruit Detection In Natural Scenes

The fruit industry is an important component of China’s agriculture.In recent years,with the development of electronic information and computer technology,automated fruit harvesting has emerged as a key research area,and efficient detection of fruits on trees is critical for achieving this goal.Traditional digital image processing detection methods have the disadvantages of complex manual feature design,poor adaptability,and generalization ability.In recent years,the accuracy and speed of deep learning object detection have made great progress.However,as a data-driven method,deep learning requires a large amount of training data in practical applications.Fruit planting usually has regional and seasonal characteristics,and fruit image acquisition needs to meet certain conditions in terms of time and space.In addition,the planting environment of fruit trees is complex,and obtaining sufficient sample image data and annotations requires high costs.Therefore,achieving accurate detection of fruit object with only a small amount of samples has become an urgent problem.This thesis focuses on the practical needs of automated fruit picking,and conducts research on few-shot object detection of fruit on trees in the natural environment of orchards.The main work is as follows:(1)A data augmentation method for few-shot object detection is proposed,which combines traditional data augmentation,DC-GAN image generation,and Mosaic data augmentation.This method uses Generative Adversarial Networks to generate completely new samples,and then combines traditional data augmentation methods and Mosaic data augmentation to further augment the generated images and real images to obtain the final extended samples.Starting from the root of the few-shot problem,this method improves the detection performance of the detection network.In the real-world detection of passion fruit,apples,and citrus,this approach improves the average precision(AP)of detection by 2.66%,6.83%,and 5.14% respectively,compared with the method without data augmentation,and is applicable to different detection networks.(2)A few-shot learning method for fruit object detection is proposed.This method is based on transfer learning strategy and addresses problems such as non-optimal solutions of sub-modules optimization due to high coupling degree in few-shot target detection network,mutual influence between classification and location tasks,insufficient learning of new class features,etc.The network is improved by introducing gradient decoupling structure,contrastive encoding classification module,optimizing training strategy,and using prototype calibration module to calibrate category scores.The experimental results show that the proposed method can improve the detection performance of fruit under few-shot conditions in fruit datasets containing passion fruit,apples,and citrus,and outperforms current mainstream few-shot object detection methods.In addition,the proposed method demonstrates good generalization performance on the public dataset PASCAL VOC and successfully improves the detection precision.