特征(语言学)
特征提取
模式识别(心理学)
计算机科学
人工智能
杂质
温度测量
数学
信号处理
条状物
电子工程
材料科学
算法
人工神经网络
工程类
生物系统
遥感
测量不确定度
作者
Hang Yin,Longpeng Ding,Zhentao Wang,Jiyong Hu,Xudong Lu,Tengda Geng,Hongfei Yang,Jinbin Li
标识
DOI:10.1109/tim.2026.3742020
摘要
Low-loss jujube harvesting requires blower power to be matched to the impurity state of jujube strips, yet reliable visual assessment remains difficult under variable illumination, dust interference, and heterogeneous mixtures of fruits and impurities. This study develops a machine-vision-based framework for six-grade jujube-strip impurity-state assessment and proposes ZPM-inspector as the perception component for adaptive fan-power regulation. ZPM-inspector combines a progressive optical enhancement pathway with an impurity-oriented multiscale feature representation to improve discrimination of elongated, fine-grained, and partially occluded impurities. A group-aware partition of 2824 images from 564 physical jujube strips/acquisition groups was established before preprocessing and augmentation to prevent information leakage. On the held-out group-disjoint test set, ZPM-inspector achieved 91.47 % accuracy, 91.46 % precision, 91.50 % recall, and 91.48 %$F1$-score, outperforming the evaluated baseline models. Relative to the consensus operational reference labels, the model obtained a mean absolute error of 0.095 grades and a root mean square error of 0.336 grades, with 99.08 % of samples predicted within one grade. In a blocked randomized field comparison with 24 trials per strategy, recognition-guided regulation reduced fruit damage by 1.69 percentage points and increased collection efficiency by 3.10 percentage points relative to fixed-power regulation. These results indicate that ZPM-inspector can provide group-validated impurity-state perception for adaptive, low-loss jujube harvesting under the evaluated conditions. The shareable data are available athttps://github.com/Y1SHAI/ZPM-Inspector
科研通智能强力驱动
Strongly Powered by AbleSci AI