Font Size: a A A

Study On Identification Of Paddy Weeds In Complex Background Based On Deep Learning

Posted on:2023-03-23Degree:MasterType:Thesis
Country:ChinaCandidate:D HuFull Text:PDF
GTID:2543306818987469Subject:Computer technology
Abstract/Summary:PDF Full Text Request
Weeds in paddy fields is an important factor restricting the increase of yield and income of rice.Correct identification and selection of appropriate herbicide are effective ways to control paddy weeds.The development of computer technology makes it possible to identification paddy weeds quickly and accurately by using machine vision technology,and provides decision-making basis for mechanized and intelligent operation of rice production management.At present,weeds identification by image processing is mostly in simple background.In this study,deep learning will be used to identification paddy weeds on the self-constructed paddy weeds image dataset with complex background taking in the paddy field.1.Paddy weeds image dataset construction and image pre-processing.Images of12 common rice weeds,such as white phosphorus sedge,polygonal lilac and multilateral water amaranth,etc.,were taken in the natural environment in the suburban area of Shanghai during the rice growing season to construct paddy weeds image dataset.The preprocessing methods of image enhancement and image segmentation are used to balance the number of images of each kind of weed and filter some interference features in the early stage of model training so as to accurately extract weed features using the recognition model during weed images training.2.Establishment of an improved model for identification of paddy weeds in complex backgrounds.An efficient and accurate recognition model,YOLOv4-Weed,is proposed based on end-to-end modeling aiming to solve the problem of segmenting weeds for the background with similar color and texture.This model performs depthwise separable convolution and inverted residual unit replaces original model standard convolution and residual unit,dense networks improve multiscale detection and add GAN improved CSPDarknet,SPP and PANet modules of the original network.The improved YOLOv4-Weed model has improved the accuracy and speed of weed identification,and also has a smaller size and better robustness.3.The result of weeds identification.Twelve paddy weeds commonly found in Shanghai is identified using the YOLOv4-Weed model.The recognition accuracy and detection speed of the model reached 94% and 60.3 fram es/s respectively,comparing with the Faster R-CNN,SSD and YOLOv4 original model,the detection accuracy improved by 5%-12% and the detection speed improved by 12.1-43.5 frames/s.The improved model has obvious advantages.The techniques,methods and models used in this study can be applied to weed identification in paddy fields with complex backgrounds,and will play an important role in weed diagnosis and control service based on machine vision,assisting intelligent farm equipment to identify target weeds and praying herbicides quantitatively.
Keywords/Search Tags:deep learning, complex background, target detection, weed identification, YOLOv4
PDF Full Text Request
Related items