| The inspection and grading of kiwifruit is a key part of post-harvest processing and an important support for commercialization and value-added.Currently,the grading of kiwifruit in China is mainly done by manual sorting,which is inefficient and subjective,and the existing sorting equipment,such as mechanical size grading and weight grading,does not allow the identification of external defects of the kiwifruit.Therefore,it is necessary to use new technologies to accurately and efficiently detect and classify kiwifruit quality by grade and to improve the commercial value of kiwifruit.The main research and conclusions of this paper are as follows:(1)To improve the quality of the acquired images,the image acquisition device was built and the light source system was optimized using the diffuse reflection principle.The camera parameters were set and tuned,and the camera was calibrated with an average reprojection error of 0.079 pixels using the Zhang Zhengyou method;the light source system was tuned and optimized using the diffuse reflection principle;a total of 1020 samples of health,wear,scar,and sunburn kiwifruit were acquired in the field and online in multiple batches,with a single image containing 1-3 samples of varying sizes,2220 images from the upper camera and 300 images from the side camera We used vernier calipers and unit area measuring paper to measure kiwifruit diameter length and maximum cross-sectional area true values.(2)To improve the accuracy of kiwifruit external defect detection,processable epidermal defect kiwifruit identification,the kiwifruit external health-defect,multi-category defect nondestructive detection method was studied.The Res Net34 model was constructed and integrated with Convolutional Block Attention Module to optimize the model training parameters,and the accuracy of this model for health-defect kiwifruit recognition was 99.5%.The comparison with Alex Net,VGG16,Inception V3,and Res Net34 verified the strong recognition ability and high stability of Res Net34+CBAM.A multi-data enhancement fusion method based on adjustable range is proposed.The YOLOv5 model was built and the model structure was improved,and the overall detection accuracy of this model for multi-category defects of kiwifruit was 97.7%.By comparing with SSD,YOLOv5 s,YOLOv7 and other models,we got the best results of YOLOv5-ours in this paper for kiwifruit external defects detection taking into account the accuracy and model size.(3)To provide a basis for processing kiwifruit by grade,a method of kiwifruit size detection was studied.A top and side dual view kiwifruit feature size detection model was constructed.Position error correction was performed according to the kiwifruit motion position state on the grading line.The predicted kiwifruit diameters,long diameters,short diameters and contour areas were obtained by the external minimum rectangle method,contour area method and coordinate system conversion.The root-mean-square errors of the predicted kiwifruit diameters,long diameters,short diameters,and contour areas were 1.49 mm,1.20 mm,0.82 mm,and 1.19 cm2,respectively,with coefficients of determination of0.89,0.86,0.84,and 0.97,thus verifying the feasibility of dimensional inspection and grading.(4)We developed software for external quality inspection and grading of kiwifruit for different grading needs such as fresh food and processing,and tested the system integration.Developed kiwifruit online defect and size inspection and grading software,and designed criteria for discriminating kiwifruit appearance defects.A total of 100 samples of each category were tested,and the test accuracy rate was 100%.A grading level of 3 was defined and the 50 samples tested as healthy kiwifruit were graded by diameter.The root mean square error of the test sample was 1.03 mm.Therefore,the real-time detection and grading of the external quality of kiwifruit was realized,and the feasibility of the test was verified. |
