Research On The Detection And Enumeration Methods Of Dense Microscopic Pests Of Large Scale And Multiple Classes

There are many different types of pests in China,which occur throughout the growth cycle of crops and are one of the main factors causing reduced grain production.How to effectively improve pest monitoring and early warning capabilities is of great significance to achieving increased grain yields,reducing the use of pesticides and improving the quality of agricultural products.At present,pest monitoring in China mainly relies on manual surveys by agricultural technicians in the field,which is time-consuming,labour-intensive,subjective and error-prone.Although computer vision-based pest image recognition technology has been developed rapidly and has solved some of the problems to a certain extent,for small target pest image recognition,there are still problems such as low recognition rate,poor robustness and weak model generalisation ability,which make it difficult to realise Engineering applications.To address the above problems,this paper focuses on three aspects of pest image enhancement,pest category imbalance and pest detection frame screening,using deep neural network technology as the theoretical basis to carry out systematic research applications for large-scale multi-category dense and tiny pest detection and counting.The following three specific areas of research and application are completed:(1)This paper discusses in detail the concepts,principles,structures,parameters and target detection networks of convolutional neural networks,and gives evaluation metrics commonly used in experiments to evaluate the superiority of the models.(2)The Pest-YOLO proposed in this paper in Chapter 3 proposes two improvements on the YOLOv4 model to address the challenges that exist in the pest dataset.The I-confidence loss(Improved Confidence Loss)proposed in this paper is a focal loss algorithm introduced in confidence loss,which can effectively address the problem of difficult sample learning.This paper also introduces a Confluence strategy to optimise the candidate frame selection for pest detection.The research in this paper proposes that Pest-YOLO’s large-scale multi-class dense and tiny pest detection and counting model has achieved impressive results.In terms of detection performance,Pest-YOLO outperforms the current mainstream SOTA detectors such as YOLOv5 s,YOLOv5m,YOLOX,DETR and TOOD.(3)This paper proposes an improved Pyramid Vision Transformer-v2-based pest detection and counting model in Chapter 4,which further optimises the large-scale multi-class dense and tiny pest detection tasks based on the tasks studied in Chapter 3.The improved model proposed in this paper is optimized for the positive and negative sample problem in model training and the multi-headed attention mechanism,and adds scale perception,space perception and task perception modules to this model,which significantly improves the performance of the model for the pest detection task without increasing the computational effort too much.In summary,this paper provides an in-depth study of the problem of detecting and counting large-scale multi-class dense and tiny pests,proposes a corresponding solution,and then conducts a full experimental validation.The proposed algorithmic model opens up new ideas for subsequent research work for reference and learning,and can also provide technical references and necessary supplements for agricultural pest control.The content of this article has been published in one paper in sci q2,and another one is pending submission.