Polarization imaging based bruise detection of nectarine by using ResNet-18 and ghost bottleneck

Bruise reduces the edible quality of nectarines and brings potential risks of cross-infection of fruit, but it is challenging to accurately detect bruised nectarines because of the color similarity between a sound and bruised area in nectarine fruit. Therefore, this research aims to develop an effective and efficient method, based on polarization imaging (PI) technology, for the detection of bruised nectarines. A total of 406 nectarines from three varieties, including 206 sound nectarines and 280 bruised nectarines made by the impact method, were used in the experiment. Polarization images of nectarines at four polarization angles (0°, 45°, 90°, and 135°) in five periods (6, 12, 24, 48, and 96 h after bruise formation), were acquired using a visible PI system. A lightweight network (ResNet-G18) that integrates ResNet-18 and ghost bottleneck was used to develop a classification model for sound and bruised nectarines based on the polarization image data. The results indicated that polarization images can effectively suppress the glare resulting from the smooth surface of nectarines, and overcome the interference of dark colors in detecting bruises. Meanwhile, the ResNet-G18 detection model achieves the best performance in both detection accuracy and efficiency, with a total classification precision of 96.21% and a true positive rate of 97.69% for bruised nectarines, and an average detection time of 17.32 ms for each fruit. Furthermore, the ResNet-G18 can recall all nectarines bruised for 12 h and above with 98.62% precision, showing excellent detection performance for early bruises. This study has demonstrated that the proposed PI technology coupled with an appropriate lightweight classification network can be effective and efficient for the detection of bruised nectarines.