Authors

Abstract

Due to their delicate and thin skin, ‘huangguan’ pears are very vulnerable to pressure and impact during picking, packing and transportation, which can cause bruising. Early detection of bruises allows for timely identification of affected fruits to reduce potential food safety risks. However, early bruises in ‘huangguan’ pears, particularly those that occur within the 30 minutes, often show no visible differences in external features compared to healthy tissue, making conventional techniques such as manual and machine vision sorting ineffective. Accordingly, a near-infrared (NIR) camera imaging technique combined with deep learning segmentation algorithm for early bruise ‘huangguan’ pears detection is proposed in this study. Firstly, a near-infrared camera imaging system is applied to collect early bruise images of ‘huangguan’ pears, and then a lightweight segmentation model based on the DeepLabV3+ architecture, referred to as MCC-DeepLabV3+ is presented. In the MCC-DeepLabV3+ model, MobileNetV2 is used as the backbone network, reducing the parameter size and enhancing deployment efficiency. Additionally, the coordinate attention (CA) mechanism is integrated into the shallow feature extraction and ASPP modules to improve the extraction of positional information across various features, minimizing the discrepancy between segmented areas and the actual bruised regions. Furthermore, a cascade feature fusion (CFF) strategy is incorporated into the encoder to reduce segmentation edge discontinuities and ensure effective multi-level semantic fusion, improving segmentation accuracy. The experimental results show that the proposed model has achieved a mIoU of 95.68%, and mPrecision of 97.58% on the self-built dataset of early bruising in ‘huangguan’ pears. Compared to benchmark models such as U-Net, SegNet, PSPNet and HRNet, the proposed model demonstrates superior segmentation performance, offering promising support for the development of nondestructive detection techniques for agricultural product quality.

Abstract in English

由于黄冠梨的表皮细腻而娇嫩，在采摘、包装和运输过程中非常容易受到压力和冲击，这可能导致瘀伤。早期发现瘀伤可以及时识别受影响的水果，有助于减少潜在的食品安全风险。然而，皇冠梨的早期瘀伤，特别是那些在30分钟内发生的瘀伤，与健康组织相比，通常在外部特征上没有明显的差异，这使得人工和机器视觉分类等传统技术的效果不佳。因此，本研究提出了一种结合深度学习分割算法的近红外(NIR)相机成像技术用于早期皇冠梨瘀伤检测。首先，采用近红外相机成像系统采集黄冠梨的早期瘀伤图像，然后，提出了一种基于DeepLabV3+架构的轻量级分割模型，称为MCC-DeepLabV3+。该模型采用MobileNetV2作为骨干网络，减少了参数大小，提高了部署效率。此外，将坐标注意（CA）机制集成到浅层特征提取和ASPP模块中，提高了不同特征之间的位置信息提取，最大限度地减少了分割区域与实际损伤区域之间的差异。在编码器中引入级联特征融合（CFF）策略，减少了分割边缘不连续，保证了有效的多级语义融合，提高了分割精度。实验结果表明，在自建的黄冠梨早期瘀伤数据集上，该模型的mIoU和mPrecision分别达到了95.68%和97.58%。与U-Net、SegNet、PSPNet和HRNet等基准模型相比，该模型显示出优越的分割性能，为水果早期无损检测技术的发展提供了有力的支持。