Study On Recognition Method Of Cucumber Fruit In Facility Greenhouses Based On Deep Learning

Cucumber is an important vegetable widely cultivated in many countries.At present,the harvesting of cucumber fruits basically depends on manpower,and the harvesting quality is uneven and inefficient.Automatic harvesting of intelligent agricultural machinery can save labor and improve efficiency,but the color of mature cucumber is similar to that of the environment,which will lead to poor recognition effect.The machine learning method using shape matching and color threshold segmentation will eliminate the background and cucumber fruit together.The method based on deep learning needs to be optimized to obtain better results.Therefore,this paper proposes a cucumber fruit recognition and detection method based on deep learning,which enhances the recognition ability of the target detection model to cucumber fruit from two aspects of color feature and network structure.The main work contents are as follows:(1)The literature on object detection technology based on machine learning and deep learning in application scenarios such as ’fruit recognition’ and ’fruit counting’ had been read.The operation principle of the target detection model based on deep learning is studied.The photos of cucumber fruit were obtained,and a cucumber fruit dataset was built and manually labeled using LabelImg software.(2)The analysis of cucumber fruit recognition algorithm based on deep learning.The Faster R-CNN,SSD,and YOLOv5 target detection models are selected,and self-built cucumber fruit data set are used for training and detection experiments to obtain the detection accuracy and detection speed of each models.And these indicators were compared with the improved results.(3)Research on YOLOv5 cucumber recognition method based on color space feature enhancement.In order to solve the problem of identifying green cucumber fruits under a near-color background,the RGB image is converted into images in other color spaces.The ReliefF weight analysis method is used to analyze the weight of information in different color spaces to distinguish between green cucumbers and green backgrounds.The Cr channel in the YCbCr color space with significant differences is selected.Before formal training,the cucumber fruit image converted to the Cr channel is sent to the YOLOv5 network model as pre-training data to enhance the model recognition ability.(4)Research on cucumber recognition method based on improved YOLOv5.CIOU＿Loss is used as the loss function of the prediction box to ensure the detection accuracy.In the process of screening candidate boxes,the Soft-NMS algorithm is used to improve the recognition accuracy of overlapping cucumber fruits.The attention mechanism is introduced to make the target detection network model allocate more weights on more important features.The results show that compared with the original YOLOv5 model,the recall rate of the proposed method increases from 44.4% to 57.60%,the m AP value increases by6.7%,the accuracy rate increases from 83.7% to 87.90%,and the FPS decreases by 7.6frames.Cr information improves the accuracy of SSD,Faster R-CNN,YOLOv5 versions and YOLOv7 by 1.51%,3.09%,3.15%,0.63%,2.43% and 1.08% respectively,which verifies the effectiveness of the improved method.