基于SIRI和CNN的苹果隐性损伤检测方法

苹果从采摘到销售过程中易发生机械损伤，需要及时剔除以避免腐烂变质。然而机械损伤早期苹果外观颜色变化不明显，通常表现为隐性损伤，检测比较困难。提出了一种基于结构光反射成像(SIRI)和卷积神经网络(CNN)的苹果隐性损伤检测方法。通过搭建SIRI系统，采集待测苹果调制的结构光图像，再利用三相位解调法提取交流分量，增强苹果隐性损伤对比度；然后利用交流分量图像制作苹果隐性损伤数据集，并使用基于CNN的语义分割网络FCN、UNet、HRNet、PSPNet、DeepLabv3+、LRASPP和SegNet训练损伤检测模型，多组试验结果表明上述模型均能有效地检测出不同情况下的苹果隐性损伤。其中HRNet模型精确率、召回率、F1值和平均交并比较高，分别为97.96%、97.52%、97.74%和97.58%,但检测速度仅为60 f/s; PSPNet模型检测速度较快，可达到217 f/s,但其检测精度略低，精确率、召回率、F1值和平均交并比分别为97.10%、94.57%、95.82%和95.90%。

Detection Method for Implicit Apple Damages Based on SIRI and CNN

During the process from harvest to sales, apples are susceptible to mechanical damage, which can have detrimental effects on their quality and lead to rotting. Detecting and removing this damage in a timely manner is crucial to prevent further deterioration. However, early-stage mechanical damage in apples often manifests as subtle color changes, making it challenging to detect. To address this issue, an apple implicit damage detection method was presented based on structured-illumination reflectance imaging (SIRI) and convolutional neural network (CNN). By building an SIRI system to acquire modulated structured light images of the measured apples, and utilizing three-phase demodulation method to extract the alternating current component, the image contrast of the apple implicit damage can be enhanced. The dataset of apple implicit damages was produced by using the images of alternating current components. Several CNN based semantic segmentation networks, including FCN, UNet, HRNet, PSPNet, DeepLabv3+, LRASPP, and SegNet were employed to train the damage detection models, respectively. Several groups of experimental results demonstrated that these models can effectively detect the apple implicit damages in different situations. In contrast, the precision (P), recall (R), F1 score, and mean intersection over union (MIoU) of the HRNet model were respectively 97.96%, 97.52%, 97.74% and 97.58%. However, its detection speed was only 60 frames per second. The PSPNet model had a faster detection speed, reaching up to 217 frames per second. However, it had slightly lower detection accuracy, with precision (P), recall (R), F1 score, and mean intersection over union (MIoU) of 97.10%, 94.57%, 95.82%, and 95.90%, respectively.