| The development of agriculture is about food security,and the development of organic agriculture is about food security.Food security guarantees the survival of China’s 1.395 billion people and 7.594 billion people worldwide.Food insecurity caused by pesticide residues and other problems will endanger national health.In the agricultural production process,precision agriculture is crucial to the challenges of productivity,environmental impact,food security and ecologically sustainable green development.Maize is the main cash crop planted in China,and the seedling stage of maize is the stage where it is most needed to be cultivated.This study takes corn seedlings as the research object,uses deep learning-based target detection as the technical means,and uses CNN visualization technology analysis and deep learning modeling to combine,aiming at corn seedling target detection from field working conditions.The research results provide new technical means and methods for the in-depth study of the identification of full-cycle corn seedlings on the multiweather and multi-angle conditions.Production control applications provide thesis foundation and technical support.Considering the huge yield of corn and the complexity of corn seedling detection,this research focuses on corn seedlings.It is of far-reaching significance to look forward to deep learning technology to help the development of agricultural robot industrialization.The main conclusions are as follows:(1)Introduce the preparatory work required for the corn seedling target detection algorithm based on field conditions.It is divided into three main parts: experimental materials,experimental equipment and software experimental materials,and field robot platform introduction and structural design.The design and construction of a field robot platform based on the openness of the agricultural production system is studied in detail.The experimental materials introduced the source and selection criteria of the experimental materials.The experimental materials were corn seedlings from the experimental field in Xiangfang District.Experimental equipment and software The experimental materials introduce the hardware equipment and software used in detail.The required experimental equipment includes: a computer workstation,three computers,a Raspberry Pi 3B +,an Intel neural computing stick Movidius 2 So C,six USB industrial digital camera.The software required for the study includes: MATLAB 2019 a,Py Charm,Label Img,and Tensor Space.The field robot platform introduction and structure design introduced the composition and function of the platform in detail.The field robot platform is the basic condition for subsequent image acquisition and testing.The field robot platform mainly includes a car body platform,a computer group,an intelligent control system,and a bi-wing vision system;the Pro / E software is used to design the car body platform and the bi-wing vision system,and the preliminary construction of the experimental platform is completed.Main basic technical parameters.(2)Realization of corn seedling target detection based on Faster R-CNN.With the help of a self-built field robot platform and data labeling tool MATLAB App Image Labeler,20,000 sample pictures were collected,marking 32,354 corn seedlings and 6,918 weeds.Faster R-CNN was built using MATLAB 2019 a,and 10 Pre-trained networks were selected using the transfer learning method to replace the CNN feature calculation part of the classic Faster R-CNN network.A Faster R-CNN with VGG19.Focusing on Faster R-CNN with VGG19 based on transfer learning processing,the average accuracy rate P is 97.71%,F1 is 97.31%,and the detection speed is 4 frames/ s.(3)From the field operation point of view,the target detection of corn seedlings under full cycle,multi-weather and multi-angle conditions is realized.The whole cycle: pictures from the end of the germination period of the caryopsis germination and soil breaking to the early stage of jointing of the corn plant were collected,and the time point of the early stage of corn planting also corresponds to the time of ridge-closing of cultivating.Corn seedlings under full cycle conditions are corn seedlings with 2 to 7 leaves.Under full cycle conditions,Faster R-CNN with VGG19 has an accuracy rate of 0.9876 for 2-5 leaf stage corn seedlings,a recall rate of 0.9646,and F1 of 0.9760;the accuracy rate for 6-7 leaf stage corn seedlings is The recall rate was 0.9497,0.9822,and F1 was0.9657.Multiple weather: Rainy,cloudy and sunny environmental conditions,including light intensity and soil type.Multi-angle: The camera’s vertical shooting angles are 75 °,30 °,and 0 °.Under multiple weather conditions,Faster R-CNN with VGG19 has a precision P of 0.9739 for rainy corn seedlings,a recall of R of 0.9609,and F1 of 0.9674;a precision of 0.95% for corn seedlings on cloudy days and a recall of 0.9843.,F1 is 0.9846;the accuracy rate of sunny corn seedlings is0.9574,the recall rate is 0.9349,and F1 is 0.9460.Under multi-angle conditions,the accuracy P of Faster R-CNN with VGG19 for 75 ° corn seedlings is 0.9820,the recall R is 0.9725,and F1 is0.9772;the accuracy P for 30 ° corn seedlings is 0.9802,and the recall R is 0.9808 and F1 are 0.9805;the accuracy rate P for 0 ° corn seedlings is 0.9676,the recall rate R is 0.9517,and F1 is 0.9596.(4)Realization of corn seedling target detection based on Mobile Net V3-SSD.During the image acquisition phase,129,792 images were obtained and 21,907 of them were labeled.Use the Tensor Flow framework to build a Mobile Net V3-SSD network,and use the 3D visualization framework Tensor Space to help understand the network layer status,coefficients,sizes,and model colors,and pruning the Mobile Net V3-SSD network.Move the pruned Mobile Net V3-SSD from the PC to the inference edge or terminal(Raspberry Pi 3B +).The focus is on the Mobile Net V3-SSD based on pruning,with an overall accuracy of 97.71%.F1 is 88+56%,and the detection speed is 18 frames / s.(5)For the initial research,the camera’s multiple angles and the entire cycle of corn seedling growth remained inadequate.The full cycle and multi-angle conditions were studied as deeply as possible.From the perspective of field conditions,the target detection of corn seedlings under full cycle and multi-angle conditions is realized.The corn seedlings under the whole cycle conditions are corn seedlings with 1 to 9 leaves.Corn seedlings under full cycle conditions are corn seedlings with 2 to 7 leaves.Multi-angle: Based on the combination of vertical rotation angle α(0 °,30 °,75 °,and 90 °)and horizontal rotation angle β(-90 °,0 °,45 °,and 90 °)to form a multi-angle of16 spatial rotation angle.Under the full cycle condition,the overall accuracy OA of Mobile Net V3-SSD with pruning for corn seedlings marked Maize3 is 0.9094;the overall accuracy OA of corn seedlings marked Maize2 is the second highest 0.8991;the overall corn seedlings marked Maize1 The second lowest accuracy OA is 0.7302;the overall accuracy OA of the weed marked Weed is at least 0.8502;under multi-angle conditions,the overall accuracy of multi-angle is divided into 4levels: OA≥95%,95%> OA≥ 90%,90%> OA≥85% and 80%> OA≥70%.In the Full-cycle condition,this research obtained the curve relationship between maize seedlings of different leaf ages and detection accuracy.In the Multi-weather condition,the detection effect of cloudy was better than rain,and the detection effect after rain was better than that of sunny.In the Multi-view condition,the OA range grades corresponding to 16 spatial rotation angles are obtained,which provides suggestions for intelligent cultivating operations. |
