Authors

Abstract

Measuring melon fruit diameter offers key insights into growth status and maturity. To overcome the limitations of manual measurement—namely high labor demands, time consumption, and large errors—this study introduces a method based on an improved Mask R-CNN algorithm. The model uses ResNet50 as the backbone and incorporates a Channel Prior Convolutional Attention (CPCA) mechanism and a bidirectional feature fusion pyramid network to enhance multi-scale feature extraction. A Self-Attention (SE) mechanism is added to the mask branch to improve segmentation accuracy. Measurement points are determined through contour segmentation, curvature analysis, and bounding rectangle fitting. A binocular camera provides depth information, and Euclidean distance is used to compute actual size. The improved algorithm achieves detection and segmentation precision of 94.2% and 92.7%, with recall rates of 94.5% and 93.6%. The method yields average relative errors of 7.1% (horizontal) and 7.6% (vertical), meeting practical agricultural needs and supporting maturity assessment.

Abstract in Chinese

测量甜瓜果实直径可以提供对生长状态和成熟度的关键见解。为了克服人工测量的局限性，即高劳动力需求、时间消耗和大误差，本研究引入了一种基于改进的Mask RCNN算法的方法。该模型使用ResNet50作为骨干，并结合了信道先验卷积注意力（CPCA）机制和双向特征融合金字塔网络，以增强多尺度特征提取。在掩码分支中添加了自注意（SE）机制以提高分割精度。通过轮廓分割、曲率分析和边界矩形拟合来确定测量点。双目相机提供深度信息，欧几里德距离用于计算实际尺寸。改进后的算法实现了94.2%和92.7%的检测和分割精度，召回率分别为94.5%和93.6%。该方法的平均相对误差为7.1%（水平）和7.6%（垂直），满足实际农业需求，支持成熟度评估。