LAUSR

BACKGROUND This paper proposed a novel method based on deeply supervised convolutional neural networks for the non-destructive detection of blueberry fruit by taking advantage of deep learning in classification tasks to improve the accuracy and efficiency in detecting quality of the blueberry fruit. We firstly collected the "Tifblue" blueberries at seven different stages of maturity (10-70 days after full bloom) and measured the pigments of the blueberry skin and the total sugar and the total acid of the pulp. We then established the skin pigment contents prediction network (SPCPN) based on the correlation between the pigments and the blueberry pictures, as well as the fruit intrinsic qualities prediction network (FIQPN) based on the correlation between the pigments and the fruit intrinsic qualities. Finally, SPCPN and FIQPN were consolidated into the blueberry quality parameters prediction network (BQPPN). RESULTS The results showed that the anthocyanin of the blueberry skin was extremely significantly correlated with the total sugar, total acid, and sugar/acid ratio of the fruit. In addition, after verification, it was indicated that for anthocyanin, chlorophyll, anthocyanin/chlorophyll ratio prediction, the SPCPN network model was found to achieve a higher R2 (RMSE) value of 0.969 (0.139), 0.955 (0.005), 0.967 (15.4), respectively. Moreover, the FIQPN network model was also able to evaluate the value of total sugar (R2 = 0.940, RMSE = 4.905), total acid(R2 = 0.930, RMSE = 2.034), and sugar/acid ratio (R2 = 0.973, RMSE = 0.580). CONCLUSION The above results indicated the potential for utilizing deep learning technology to predict the quality indicators of blueberry before harvesting. This article is protected by copyright. All rights reserved.