*3.3. Results of the Defective Apple Grading*

In order to verify the detection effect of the proposed method of defective apple grading, 68 first-class apples, 64 second-class apples and 62 third-class apples were purchased in a supermarket and selected for testing referring to the grading standard of apples in Beijing. The experimental results are shown in Table 4 and Figure 10.

**Table 4.** The average of detection results of three grades.


As shown in Table 4, the precision and recall in the three grades of apples were above 90.63%, and the overall precision and recall were 94.30% and 94.33%, respectively. The detection accuracy of apples was 92.42%, and the F1 value was 94.31%. Among the three grades, the apple grade identification with the highest precision was the first-class apples (95.59%), and the ones with lowest precision were second-class apples (92.06%). Due to the error that would occur when the area values of defect were close to the junction of two adjacent grades, first class and third class might be misclassified into second class, and second-class could also be misclassified into first class and third class. Meanwhile, if the defect was located at the edge of apple, it was sometimes incorrectly detected as the background by the object-detection algorithm and semantic segmentation model, which would also lead to the reduction of classification accuracy.
