Research And Application Of YOLOv4-Tiny-CBAM Based Weed Identification In Peanut Seedling Stage

Peanut is one of the common oilseed crops in my country.Weeds in the seedling stage seriously endanger the yield and quality of peanuts and restrict the growth of peanuts.The rapid development of artificial intelligence has greatly improved the efficiency of weeding robots,but factors such as similar leaves of companion weeds at the peanut seedling stage,complex field background,variable light intensity and target shading have increased the difficulty of research.Existing machine weeding mostly uses filming for recognition due to the large mounted model and slow computing speed.Therefore,researching an algorithm that enables real-time,accurate identification of peanut seedlings and weeds under field conditions is important for the development and application of weeding robots.The main research work and innovations in this paper contain three parts as follows.(1)Constructing a peanut seedling weed dataset.In this thesis,we studied peanut seedlings and weeds in a field under natural lighting and complex background.Label Img annotation tool was used to annotate the peanut seedling weed images to create a peanut seedling weed dataset containing 6928 images of peanut and six common weeds in peanut fields,including Amaranthus concolor,gibberellic acid,chrysanthemum,common milkweed,small thistle and duck-plantar grass.The dataset was divided according to the ratio of training set,validation set and test set = 6:2:2 for model training and testing.(2)A YOLOv4-Tiny-CBAM-based peanut seedling weed recognition and detection algorithm is proposed.In order to solve the problem of detecting peanut seedling weeds with small targets,large numbers and complex backgrounds,and to improve the detection effect,this paper introduces the convolutional attention mechanism into YOLOv4-Tiny,proposes a peanut seedling weed recognition and detection algorithm based on YOLOv4-Tiny-CBAM,and compares the effects of fusing different attention mechanisms on the recognition accuracy and recognition speed of YOLOv4-Tiny The effect of incorporating different attention mechanisms on the recognition accuracy and recognition speed of YOLOv4-Tiny.The comparison tests with Faster R-CNN algorithm and SSD algorithm showed that the average recognition accuracy of the proposed algorithm for weeds reached 90.39%,and the single detection time was 44.79 ms,which proved that the constructed peanut seedling weed recognition model had better effect,with good recognition effect and generalization ability,and could better realize the weed recognition in peanut seedling stage.(3)A mobile terminal for peanut weed identification at the seedling stage was designed and developed.The mobile terminal uses the NCNN open source framework to encapsulate the model and return recognition results by calling up the camera or selecting videos from a photo album and sending them to the server.The terminal also contains a weed control encyclopaedia with detailed information and control measures for all types of weeds,including peanut.After repeated tests on videos collected from natural peanut fields with multiple video streams in different contexts,the mobile terminal was able to basically achieve good detection of peanut seedlings and weeds,providing a reference for piggybacking onto weed control robots and providing important support for intelligent research and application of weed identification at the peanut seedling stage in China.