基于特征融合与半监督协同训练随机森林的包装袋缺陷识别

Recognition of Packaging Bag Defects Based on Feature Fusion and Semi-supervised Co-training Random Forest

针对工业生产线中人工检测包装袋缺陷效率低的问题,提出一种基于特征融合与半监督协同训练随机森林的包装袋缺陷识别算法。首先,提取包装袋的颜色和边缘特征,并采用早融合得到融合特征,以增加特征区分度;然后,采用随机森林(RF)和支持向量机(SVM)半监督协同训练,从无标签的样本中选取最有代表性的样本增强样本集,使用增强样本集对RF进行训练,使分类器的识别准确率提高;最后,经过实验验证本文方法的有效性:较颜色和边缘特征,融合特征识别准确率分别提高8.75%和1.24%;较半监督协同训练,特征融合的半监督协同训练识别准确率平均提高3.75%。

A packaging bag defect recognition algorithm based on feature fusion and semi-supervised collaborative training of random forests is proposed.It solves the problem of low manual detection efficiency of packaging bags in the industrial production line.Firstly,the color and edge feature of the package are extracted and fused to obtain fusion characteristics.The fusion feature enhances the characteristic region.Then,the random forest(RF)and support vector machines(SVM)are used for semi-supervision collaborative training.It selects the most representative sample enhancement sample set from the label sample,and uses the enhanced sample set to train the Random Forest classifier.Enhanced sample sets improve the identification accuracy of the random forest classifier.Finally,the effectiveness of this method increases by verified by experiments.Compared with color and edge features,the fusion feature recognition accuracy is 8.75%and 1.24%,respectively.Compared with semi-supervision collaborative training,the accuracy rate of packaging bag defects under the semi-supervising synergistic training method of feature fusion increases by 3.75%.