Detection of bruising in pear with varying bruising degrees and formation times by using SIRI technique combining with texture feature-based LS-SVM and ResNet-18-based CNN model

Bruising negatively affects the appearance and sensory quality of fruit, weakening their commercial value. Optical non-destructive testing of bruised fruit is an important research topic. However, the degree and formation time of bruising have a significant impact on detection accuracy, which poses a challenge for accurately detecting bruised fruit. This study attempted to use structured-illumination reflectance imaging (SIRI) to detect bruised pears with different degrees and formation times. Structured-illumination images of pears with three bruising levels (S1, S2 and S3) and five bruising formation times (0, 6, 24, 48 and 72 hours) were collected. Two machine learning models including texture feature-based least squares support vector machine (LS-SVM) and ResNet-18-based convolutional neural network (CNN) were established using different input images (DC, AC, RT, DC-AC and DC-AC-RT). Study indicated that AC and RT images have depth resolution capability, achieving good results for detecting pear bruises within 24 hours combining with two types of models, showing that the powerful detection capability of SIRI technology for early bruising. Compared to LS-SVM model, the ResNet-18-based CNN model obtained better detection performance, and its detection accuracy was not significantly affected by the bruising degree and formation time. For all samples, ResNet-18-based CNN model, combined with AC and RT images, achieved an overall detection accuracy of 92.5 % and 94.3 %, respectively. This study provided a valuable solution for accurate identification of bruised pears with varying bruising degrees and times.