Abstract:

【Objective】Camellia oleifera, one of the four major woody oilseeds in the world, has a comprehensive utilization value. As a key rural industry in southern China, it plays an important role in the development of the agricultural economy. In recent years, with a vigorous support from the national government at all levels, the C. oleifera planting area has been expanding annually. Therefore, it is important to obtain information on the shapes of C. oleifera fruits quickly and accurately to determine the size distribution, guide fruit grading, and calculate rapid yield measurements. This paper proposes a batch extraction method for C. oleifera fruit shape feature parameters based on camera photography. 【Method】First, the picked C. oleifera fruits were placed on a background plate with a scale, using a camera to quickly obtain images of C. oleifera fruits and a correction of the geometric distortions produced by photography. Then, the Mask R-CNN model was used for a fast detection and counting of the C. oleifera fruits in the image. According to the generated mask, the number of pixels of the feature parameters (length of the long axis and short axis, area and perimeter) of the camellia fruits was counted using the elliptic fitting method. Finally, the image pixel size calculated with the help of the background plate scale was used to obtain the key parameters of the oil tea fruits, while the accuracy was verified using the actual measured values. 【Result】The average recognition accuracy and recall of the mask R-CNN model were 99.55% and 91.19%, respectively, and the F-measure value was 95.22%, showing that the method was suitable for describing C. oleifera fruit shape parameters statistically. The estimated accuracy of the area of the C. oleifera fruits was the highest with the coefficient of determination (R²), mean absolute error (MAE), and root mean square error (RMSE) of 0.999 0, 10.75 mm², and 14.88 mm², followed by the perimeter and length of the long axis, and the lowest accuracy of length of the the short axis with R², MAE, and RMSE of 0.864 7, 3.15 mm, and 3.74 mm, respectively. 【Conclusion】The method achieved a rapid and accurate counting of C. oleifera fruits shape feature parameters after picking and batch extraction. Thus, the technique can rapidly and accurately detect a large number of fruit feature parameters and provide a scientific basis for the C. oleifera fruit grading and rapid yield measurement.

Key words: close-range photogrammetry, shape feature extraction, image processing, convolutional neural network, Camellia oleifera fruit