基于特征融合的多尺度窗口产品外观检测方法

Product Appearance Detection Based on Feature Fusion and Multi - Scale Window

摘要:

为提高产品外观质量的检测精度和实时性,提出一种基于特征融合的多尺度滑动窗口机器视觉检测方法；在训练阶段,首先提取图像的HOG特征和Lab颜色特征,并采用典型相关分析法(CCA)进行特征融合；接下来,采用支持向量机(SVM)对融合的特征进行训练,生成分类器；在检测阶段,产品外观不同区域对精度的要求不同,为提高检测效率,生成不同尺度的滑动窗口,在每个窗口中都进行图像的特征提取与特征融合；最后,对采集的图像序列进行匹配,实现产品外观划痕的实时检测；实验中,选取不同的特征提取方法进行对比,并分别生成大小不同的滑动窗口,通过分析实验结果,结合检测时间与精度,确定各个区域的窗口尺度；实验表明,与传统的检测方法相比,所提方法在检测精度和实时性上具有显著提高。

Abstract:

In order to improve the detection accuracy and real-time performance of the product appearance quality, we present a method of machine vision on the basis of feature fusion and multi-scale sliding window. In the training phase, in the first part of the paper, the HOG feature and Lab color feature of the image are extracted, and then the feature are fused using the classical correlation analysis(CCA). Next, support vector machine(SVM) is used to train the fused features to generate a classifier. In the detection phase, the accuracy requirements of different regions of the product are different. In order to improve the detection efficiency, the sliding window of different scales is generated. In each window, the feature extraction and feature fusion are carried out. Finally, the acquisition of the image sequence to match, to achieve real-time detection of product appearance scratches. In the experiment, different feature extraction methods are selected to compare and generate sliding windows with different sizes respectively. By analyzing the experimental results and combining the detection time and precision, the window scales of each region are determined. Experiments show that compared with the traditional detection method, the proposed method has a significant improvement in detection accuracy and real-time.