Design And Research Of Weed Recognition System Based On GAN

Rice is an important food crop in China.Under the background of smart farming and precision agriculture,selective weed control will become the future direction of precision agriculture,and weed recognition based on machine vision is the base of it,so the research work of paddy field weed recognition technology is of great significance.Malignant weeds of paddy field were selected as target in this research,and a brand new paddy field weed recognition model was designed based on the Generative Adversarial Network(GAN)and Mask R-CNN.Framework of the weed recognition model constructed by image enhancement model based on GAN and weed recognition model based on Mask R-CNN.In this paper,we use CCD camera to collect weed image samples,and a Paddy Field Malignant Weed dataset(PFMW)was constructed.Experiments of PFMW dataset were done based on traditional image recognition methods and deep learning models.The main work and conclusions of this paper are as follows:(1)Construction and preprocessing of PFMW dataset.In this paper,paddy experimental field in South China Agricultural University was selected as the experimental site.Malignant weed species of PFMW dataset include: bidens,goose starwort,gomphrena,sprangle,eclipta and wedelia.Experiments have finished in this paper include edge detection experiments,image segmentation experiments and controlled experiments between traditional methods and deep learning methods.In the experiment of traditional image feature extraction method,this paper completed the LBP and HOG feature extraction of PFMW dataset and built the traditional machine learning models.The HOG-XGB model achieved the highest precision among all the traditional machine models.The F1-Score in each category of weeds was 0.8272,0.6667,0.3949,0.7037,0.8808,0.8309,respectively,and the average F1-Score was 0.7281.In the deep learning models comparison experiments,VGG16 model achieved the highest precision among all the deep learning models,and its F1-Score in each category of weeds were 0.9576,0.9316,0.9556,0.9557,0.9238,0.9925,and the average F1-Score was 0.9539.(2)In optimizer experiments,the VGG16-SGD model achieved the highest precision,and its F1-Score in each category of weeds were 0.9870,0.9746,0.9655,0.9677,0.9896,0.9825,respectively,and its average F1-Score is 0.9777.In the dataset category sample quantity equilibrium experiments,the accuracy of the VGG16 model trained by the balanced dataset is 0.9006,while accuracy of the models trained by the 16.7%,33.3%,and 66.6% category imbalanced dataset were 0.8889,0.8664,and 0.8452,respectively.(3)Construction and experimental work of a paddy field weed image enhancement model based on GAN.To solve the problem of insufficient image samples of the dataset and the inability of traditional image enhancement methods,the paper proposed image enhancement model based on GAN.Compared with the traditional image enhancement method,the model was able to target the foreground region of sample image and the paper designed experiments of image quality evaluation at the same time.Compared with the original image,the entropy of enhancement images of bidens,goose starwort,gomphrena,sprangle,eclipta and wedelia have 1.50%,2.56%,3.78%,4.92%,2.14%,2.33% increment.In the experiments of optimal ratio of enhancement set and original set,the F1-Score of VGG16 model is 0.9056,0.9117,0.9374,0.9562,0.9205 and 0.8313 respectively when the proportion of enhancement set is 0%,10%,20%,30%,40%,50%.The experiments showed that the optimal ratio between the enhancement set and original set is 30%.(4)Construction and experimental work of paddy field weed image recognition model based on Mask R-CNN.This paper has completed the annotation work of PFMW and adjusted the architecture of original Mask R-CNN model to fit it.The average accuracy of the model is 0.9018,the average accuracy of foreground region recognition is 0.8873,and the average accuracy of instance segmentation is 0.8708.Experiments in this paper have shown that the advanced technologies such as machine vision are able to locate and recognize malignant weeds in paddy fields accurately,which is of great significance in promoting fine farming in rice fields and variable spraying of herbicide.The selective weed control research in the farming production process has high economic and social value.