Research On Field Weed Segmentation And Location And Multi Classification Recognition Technology Based On Deep Learning

Associated weeds seriously affect the yield and quality of crops.Therefore,inhibiting weed growth is one of the main operations of agricultural production.Precise mechanical weeding or targeted herbicide spraying by robot is a feasible choice to reduce the impact of agricultural chemicals on the environment.In complex environment,it is difficult to accurately obtain the contour area and regional location of weeds in crop field.At the same time,in addition to the detection of weeds in complex environments,the classification and identification of specific weed species is also essential in order to apply targeted control mechanisms to weeds for accurate spraying(such as appropriate herbicides and correct doses).In this paper,using the advantages of self-learning feature ability and strong generalization of deep learning,we carry out the research on field weed segmentation and location and multi classification recognition technology based on deep learning.The main research results are as follows:(1)In order to improve the identification and location accuracy of small plant weeds and densely distributed weeds in the complex environment,an improved weed segmentation algorithm based on UNet network is proposed.Firstly,in the feature extraction process,a set of dilated convolution sequences is used to expand the receptive field of the feature map and improve the segmentation and positioning ability of small plants and weeds;Secondly,adding boundary weighted loss function to the loss function to refine the segmentation boundary can better solve the problem of overlapping crops and weeds,and obtain the regional position and contour area of weeds;Finally,experiments were carried out using self built corn data set and sugar beet data set by using a variety of deep learning networks.The results show that,the mean intersection over union(MIOU)and category mean pixel accuracy(MPA)of the improved UNET network are higher than other algorithms,which verifies the feasibility and robustness of this research method.(2)In order to facilitate the application of targeted control mechanism,it is necessary to classify and identify specific weeds.In this paper,the multi classification identification of weeds is realized by improving Dense Net network,and the herbicide types and doses suitable for each kind of weeds are determined.Firstly,after each convolution layer,efficient channel attention(ECA)mechanism is introduced to strengthen weed features and inhibit the extraction of background features,so as to improve the accuracy of weed recognition.In addition,Drop Block regularization is added after each Dense Block block to ensure the generalization ability of the model;Finally,experiments were carried out on maize seedlings and six types of associated weeds.The results showed that the comprehensive performance of the improved densenet network was better than the comparison network,and realized the efficient and accurate classification and identification of weeds in complex environment.(3)Based on the above research results,through the use of tools such as Py Charm,Qt Designer and Py QT5,the weed segmentation and positioning and multi classification recognition software is developed and designed,including research and automatic modes.The research mode includes training network weight,weed segmentation and positioning and weed multi classification recognition functions.The results show that the software can accurately and quickly identify and locate weeds in the field,and the effect is remarkable.