Recognition Methods For Typical Fruits And Vegetables Based On Deep Learning Networks

As an emerging branch of artificial intelligence,deep learning technology has unique advantages in image recognition.The related research results show that the recognition of typical fruits and vegetables based on deep learning has achieved notable results.The new progress has important implications for the development and application of intelligent detection devices and harvesting robots.The research on deep learning is still in its infancy.However,the application of deep learning to the development of intelligent agriculture has a broad prospect,it is believed that deep learning will be a research hotspot in the next few years.Based on the project requirements of a research center,this paper deeply investigates the image detection and recognition status of typical fruits and vegetables such as lettuce,rape,tomato,etc.,and selects several typical deep neural network algorithms in deep learning--YOLO series of algorithms.According to different image characteristics of fruits and vegetables,the series of algorithms for the different degrees of improvement,the above several typical fruits and vegetables on the valid identification.The main research contents are as follows:(1)A study based on YOLOv3-tiny is conducted to investigate the quality identification of lettuce in assembly line.A simple and efficient detection model is required for rapid and accurate identification of lettuce quality.YOLOv3-tiny is considered to have such advantages.Thus,the established lettuce dataset is trained by YOLOv3-tiny.Adjust the training parameters,and get the appropriate training model.The trained model is verified by the lettuce test set.The test results show that the identification accuracy was 95.6% and the detection time was 5.7ms,so YOLOv3-tiny meets the requirements for distinguishing lettuce quality.(2)A study based on D-YOLOv3 is conducted to investigate the image recognition of leaf vegetable seedlings in greenhouse.Aiming at the problem of YOLOv3's poor detection ability and long detection time for small seedling targets in images,the model is improved by using dense connection to replace the residual structure in original network and optimizing the multi-scale detection structure and loss function.The effectiveness of the original and improved YOLOv3 models is compared by tests.The results show that the detection accuracy of D-YOLOv3 was improved by 9.4% and its detection time was reduced by about 4ms.(3)A study based on IMS-YOLO is conducted to investigate the identification of tomato fruit in greenhouse.The goal of detecting and identifying tomato fruit in the complex environment of greenhouse is to assist the robots for harvesting,so it has higher requirements for the detection speed and accuracy of model.By analyzing the existing YOLO network structure,a novel network structure is reorganized and improved.The test results show that the proposed IMS-YOLO structure provides better comprehensive performance.The model is then integrated into the harvesting robot system.The harvesting test in greenhouse shows that the model has a high recognition rate.In conclusion,different recognition algorithms are provided for different fruits and vegetables,and have proven effective by tests.The findings offer a new research approach for recognition of fruits and vegetables,and also lay a technical foundation for intelligent detection devices and harvesting robots.