面向图像的有效目标区域提取方法

Extraction Method for Key Object Region in Images

针对图像标注、目标识别等实际应用中图像的前景目标定位不够准确的问题,提出了一种图像中的有效目标区域提取方法。该方法以提取图像的前景目标为目的,将目标区域提取问题转化为二分类问题,实现了对图像中有效目标区域的提取,主要包括4个步骤:利用选择性搜索算法生成图像中的候选目标区域;通过对像素值的差值化处理来进行图像区域的特征增强;基于深度学习实现对候选目标区域进行分类;区域选择与融合。在MSCOCO数据集上进行实验,结果表明,该方法在保证较高召回率的基础上,达到了比现有多种算法更加准确的目标区域定位结果。

To improve the performance of object region extraction, which is usually used in some practical applications such as image annotation and object detection, we present a new approach to the key object region extraction from images. The method ignores the category information of various objects and transforms the task into a binary classification problem. Our method includes four steps: First, a selective search algorithm is used to generate candidate object regions in the image;Second, the feature enhancement of image regions is performed by difference processing of the pixel values;Then, we classify the candidate object regions based on deep learning;Finally, region selection and fusion are finished. Experimental results over MSCOCO dataset demonstrate that the proposed approach achieves more accurate results than several related methods with high recall rates.