Research On Transformer-based Image Recognition Method For Forestry Pests

In recent years,image recognition technology has been rapidly developed and is widely used in agriculture,greatly reducing the impact of pests and diseases on crops.Today,the following problems exist in the field of agricultural pest recognition:there is a lack of available datasets for the field of forestry pests,and some of the datasets collected based on laboratory environments cannot be adapted to the needs of applications in complex environments;on the other hand,there are limitations in the extraction of fine-grained features by models,which leads to the recognition performance of small target pests not meeting the realistic needs.To address the above problems,this paper conducts research on a deep learning-based approach to forestry pest classification and detection,with the following main work:（1）Construction of dataset.A forestry pest dataset（FPD）was created.The FPD-Classification dataset for classification and the FPD-Detection dataset for the detection task were selected with data enhancement such as rotation,flipping,and adding noise for sample expansion and annotation information.（2）An improved self-supervised Mo BY pest classification model is proposed.The Mo BY lacks the extraction of fine-grained features of the image,which leads to the low performance of the model on fine-grained targets.To address this problem,this paper proposes a lightweight Pixel-decoder module based on the pixel reconstruction module,and constructs a new network model named Pixel＿Mo BY based on the Mo BY model.The effectiveness of the model improvement is verified through experiments,and the best network parameters are found for the model through ablation experiments.The accuracy of the Pixel＿Mo BY model is experimentally validated to 74.41%on the FPD-Classification dataset,which is a 3.32%improvement over the Mo BY model,and a 2.78%improvement on the downstream small target detection task.（3）An improved DETR object detection model is proposed.In order to balance computational effort and performance,an improved DETR network model is designed by adding a lightweight feature extraction branch to improve the detection performance of the model on targets at all scales.The model is named DW＿DETR,and the effectiveness of the model improvement is verified through experiments.The DW＿DETR model achieves an accuracy improvement of 8.4%and 6.1%on m AP andm AP_small respectively compared to the original model on the FPD-Detection dataset.