Research On Improved CNN Algorithm And Its Application In Recognition Of Weeds And Beets Under Night Background

Weeds in the field are harmful to the growth of crops such as sugar beet,and will compete with crops for sunlight and nutrients.If weeds are not removed in time,they will eventually lead to crop yield reduction.Therefore,timely and accurate weed removal in the field gradually becomes the focus of people's attention,and the premise of weeding requires accurate identification.In recent years,with the concept of precision agriculture being put forward,artificial and intelligent identification methods are generally used for weeding to ensure the healthy growth of sugar beet and other crops.Artificial identification mainly relies on subjective experience and requires a lot of labor.Intelligent weed identification methods are mainly divided into remote sensing identification and machine vision identification.Remote sensing can identify weeds by collecting spectral information of weeds,but it is not ideal for weeds with low growth density,therefore,at present,the mainstream method is still machine vision recognition.However,the method based on machine vision can only extract shallow features such as texture and color.In the actual complex field environment,the extracted feature information is not clear due to factors such as light transformation,thus affecting the recognition accuracy.As a kind of deep learning,convolutional neural networks can effectively avoid manual feature extraction that relies on prior knowledge,which is better than traditional intelligent detection methods.However,the existing unimproved deep learning model also has low identification accuracy when detecting weeds and crops in a complex background.At the same time,weeding is a labor-intensive and time-consuming operation.In order to improve weeding efficiency and reduce labor costs,weeds need to be identified not only when the light conditions are good during the day,but also in the absence of light at night.In this regard,this paper takes the photos of beet seedlings and weeds in the actual field environment lacking light at night as the research object.And then a deep fusion algorithm based on convolutional neural network is used to perform multi-modal fusion of weed visible and near infrared images.Finally,a improved detection algorithm is combined to realize the accurate detection of weeds.The main contents are as follows:1.In view of the problem that the collected visible pictures may contain noise and are susceptible to the influence of illumination changes,losing important details such as the shape and color of weeds and crops,a multi-modal image fusion based on deep learning is proposed.Based on the deep fusion algorithm proposed by Li et al.,the visible and near-infrared images of weeds are fused at pixel level to fully demonstrate the feature display ability of multi-mode fusion images.2.Due to image translation,rotation and other transformations,the same target may be deformed in the image,resulting in the model is not sensitive to the edge feature information of weeds,and the poor adaptability of geometric deformation weakens the fine-grained feature extraction ability of the model.In order to solve this problem,this paper proposes deformable convolution,Detnet-59 and other methods to make the model have the ability of adaptive fitting the geometry of beets and weeds.3.In the field environment,the problem of overlap and occlusion between beet seedlings and weeds may lead to a high rate of false detections and missing detections of the model,a guided generation area suggestion frame algorithm and an improved non-maximum suppression are adopted to solve it.These methods can optimize the network,so as to improve its learning ability for difficult samples.The results show that our Defor-DetNet-SoftNMS algorithm proposed can significantly improve the weed recognition performance under the complex background,and the mean average precision can reach 91.7% on the test set of multi-mode fusion image data,compared with the methods based on classic R-FCN and YOLO V2.4.After the implementation of the optimal model training,a weed and beet seedling identification system is designed.The intelligent weed detection system is designed and developed by PyQt5 and Qt Designer software,and the GUI interface of each link is also designed respectively,as well as the interface display after different operation steps.In general,the multi-modal depth image fusion method based on convolutional neural network and the improved weed and beet recognition algorithm proposed in this paper can effectively detect weeds and beets in the complex field background with lack of light at night,and also has a high recognition accuracy.Finally,the intelligent weed detection system is designed and developed by relevant software,which can lay a theoretical basis for the development of the following all-weather automatic weeding robots,and has theoretical significance and practical application value.