【摘要】 为了提高机器视觉系统的图像分割精度,提出了一种以小波变换为分类特征的多神经网络(Multi-Neural Network,MNN)图像分割算法,该算法包括小波特征提取、MNN区域划分和MNN分类3个阶段。其中,小波特征提取根据小波变换的各层图像扩展得到分类特征;MNN区域划分将初分割边界附近区域分为训练样本区域和待分类区域(待分区域),并用多边形拟合算法将待分区域划分为多个局部待分区域;MNN分类将每个局部待分区域的像素用区域内的神经网络分类器进行分类,确定目标像素和背景像素,将目标像素合并后再进行一定的后处理即可得到分割结果。以轴承表面缺陷检测系统采集的轴承缺陷图像为对象,对MNN算法和阈值分割算法进行了对比试验,结果显示MNN算法的像素数量误差(Pixel Error,PE)相比阈值分割算法降低了75%,分割精度显著提高。更多还原

【Abstract】 To improve the image segmentation accuracy of machine vision system,a Multi-Neural Network( MNN) segmentation algorithm based on wavelet transform is proposed. The algorithm consists of three phases: multi-neural network region division,feature extraction and classification. The training region and the region to be classified near the boundary of the initial segmentation are divided into several small regions by the polygon fitting algorithm. The feature extraction is achieved by the extension of images which are the result of wavelet transform of the original image. The neural network classifier classifies the pixels in the region to be classified to target pixels and background pixels,and to obtain the segmentation results,some post-processing are performed. The multi-neural network algorithm and the threshold segmentation algorithm are compared with the segmentation accuracy of bearing defect image. The result shows that the Pixel Error( PE) of the multi-neural network algorithm is reduced by 75 %than the threshold segmentation algorithm. Segmentation accuracy is improved significantly.更多还原