Predicting soluble solids content in “Fuji” apples of different ripening stages based on multiple information fusion

Abstract

Soluble solids content (SSC) is one of the important components for the assessment of fruit quality. Near-infrared (NIR) spectroscopy, multispectral and hyperspectral imaging techniques are widely used methods for the SSC estimation. However, the high prices are unaffordable for small commercial orchards. The objective of this work is to present a computer vision system for the SSC estimation of “Fuji” apples in different ripening stages. Multiple information of the color channels (channel a* and channel b*) in L*a*b* color space translated from RGB images were applied to extract color features. Stacked autoencoder (SAE) algorithm was used to extract color features in pixel-level, and the trained parameters were used as the initial parameters of the prediction model by using back propagation neural network (BPNN) algorithm. The results show that the prediction accuracy of the SAE-BPNN model based on color features in pixel-level is higher than that of the BPNN model based on pure color features in feature-level. Moreover, the model with fusion information of both color channels as inputs can get better prediction results than that of single channel. The SAE-BPNN model with SAE structure of 270-120-30 yielded the best result with Rp2 = 0.5953 and RMSEP = 0.8856%. This research indicates that RGB imaging technology based on multiple information has potential for the SSC prediction, which provides reference values for fruit harvesting and non-destructive evaluation of fruit quality.