Evaluation Of Tomato Growth Status In Solar Greenhouse Based On Digital Image Processing Technology

The detection of tomato growth status in solar greenhouse is a very important part in the process of tomato production.At present,the main method to realize the detection of tomato growth state in solar greenhouse is manual detection.This detection method not only requires regular records by personnel with rich disease knowledge and experience,but also increases the labor cost.Moreover,it is unable to carry out effective detection and recording in the dark light environment such as at night,resulting in the inability to quickly respond to diseases and other emergencies with strong timeliness.Therefore,this paper combines the digital image processing technology and the deep learning algorithm of target detection,and puts forward a research on the detection of tomato growth state in Solar Greenhouse Based on the digital image processing technology.Aiming at the problem of difficult recognition in complex light environment,the target detection model is improved,and the information transmission mode of the model is optimized.Based on the model,a fast recognition software of tomato image at night is developed.The main work and conclusions are as follows:(1)By analyzing the working principle of the YOLOv5 target detection model,the network architecture of YOLOv5 is split,and an improved night tomato target detection model of YOLOv5 is established.The anchor calculation function is introduced to machine learning,which improves the efficiency of capturing small target eigenvalues,reduces the deviation of the real detection frame,and improves the recognition accuracy of the occluded object;The ciou loss function is used to replace the original giou loss function,and the center point distance measurement is added,which can directly minimize the distance between the two target frames and improve the convergence speed.The modified model is compared with the traditional YOLOv5 target detection model and the fast r-cnn model under the tomato image data in the same environment.The experimental results show that the detection time of the improved model is 6.7 times that of the traditional CNN model,and the intersection and comparison of the improved YOLOv5 is 3.5% lower,indicating that the improved YOLOv5 has a more prominent ability to extract different tomato fruit feature information and distinguish background and target information,and the robustness and detection accuracy of the improved YOLOv5 target detection model have been significantly improved.(2)The data enhancement mode is improved,and the active learning strategy is introduced to replace the original mosaic data to enhance the image data of tomato surface defects,which solves the problem of incompatible model size and recognition accuracy,reduces the manual annotation cost and improves the annotation efficiency;The original feature pyramid network structure is replaced by af-fpn network structure,and the attention mechanism and feature enhancement module in the parallel mode are integrated to reduce the information loss in the process of feature map generation and improve the operation speed.The optimized YOLOv5 target detection model is compared with the traditional model.The experimental results show that the recognition accuracy and speed of the optimized model are improved by 3% and 1.6% respectively.(3)Based on the architecture of the improved YOLOv5 target detection model,a mobile phone application for tomato recognition at night is developed.The adaptability of the network model architecture is adjusted to enable the software to adapt to the Android system on the mobile phone.The mobile phone application made by the improved YOLOv5 model is compared with the mobile phone application developed by the traditional model.The test results show that the mobile phone application of the improved target detection model has a single fruit recognition accuracy of 100% for different mature conditions,and a multi fruit recognition accuracy of 96% for occluded conditions.Compared with the mobile phone application of the traditional target detection model,the recognition accuracy and speed are improved.