Application Research On An Online Intelligent Recognition System For Major Pests In Rice Fields

Rice is one of the main staple crops in China,and its yield and quality are important indicators for evaluating the development of agriculture.Rice pests are a crucial factor that restricts the development of the rice industry.However,under the current pest control system,the difficulty and inefficiency of pest monitoring and early warning urgently need to be addressed.With the development of smart agriculture,the establishment of an intelligent monitoring and early warning system for rice pests has gradually become an important part of achieving green prevention and control,with the primary task being the precise identification of target pests.Currently,machine learning algorithms and convolutional neural networks under the branch of deep learning,combined with computer vision technology,are widely used in the field of pest identification,relying on intelligent hardware devices such as insect condition monitoring lamps and remote monitoring equipment for rice planthoppers,and gradually transitioning towards online identification on cloud platforms,driven by big data as the core.This article focuses on two main pests in Shandong Province,the rice stem borer and the rice planthopper,and proposes an improved Grab Cut segmentation algorithm.An insect recognition model based on Mask R-CNN and YOLO＿v4 was constructed,and the specific research contents are as follows:(1)The insect sample image collection work has been completed,and a target dataset has been constructed.Firstly,the photo-type intelligent insect monitoring lamp and portable rice planthopper remote monitoring device were used as trapping and monitoring equipment to carry out image sample collection work.Secondly,the improved Haar feature algorithm was used to obtain individual insect image samples,and conventional data augmentation methods were used to expand the samples.The training dataset was supplemented with a diversity of information technology such as GAN generation network models and Python web crawlers.Finally,an improved Grab Cut segmentation algorithm was proposed,combining extreme point features and HSV color space.The algorithm was optimized and selected for complex background images,and a rice field pest dataset was compiled.(2)A recognition model of Chilo suppressalis based on Mask R-CNN was constructed and optimized.Based on the rice stem borer image dataset,the recognition network was reconstructed and optimized from two aspects: multi-level residual connection and ECA attention module.The improved model was evaluated using commonly used evaluation metrics and model comparison tests.The average precision AP of the improved model was92.71%,the recall rate R converged to 89.28%,and the F1 score was 0.9096.Finally,the actual recognition effect of the deployed model was compared and analyzed from two aspects:simulated laboratory environment and real field environment.(3)A rice planthopper identification model based on YOLO＿v4 was constructed and optimized.The recognition network was reconstructed and optimized from three aspects: the normalization method of BRN layer,the CBAM attention module,and the Focal Loss loss function,based on the rice planthopper image dataset.The improved model was evaluated using common evaluation indicators and compared with other models.The average precision(AP)of the improved model was 86.65%,the recall rate(R)was 88.85%,and the F1 score was 0.8774.Finally,the recognition accuracy and speed of the model were tested based on a cloud platform virtual environment.