旋极信息利用视觉显著性评价提出景物特征提取方案

Feature extraction is a concept in computer vision and image processing, which refers to using computers to extract image information and determine whether each image point belongs to an image feature.

The result of feature extraction is to divide the points on the image into different subsets, which often belong to isolated points, continuous curves or continuous regions. The quality of features has a crucial impact on generalization performance. Commonly used scene feature extraction methods usually include point, line, edge and other features for extraction, but these feature extraction methods usually only focus on local areas and have poor anti-interference ability.

To this end, Xuanji Information applied for an invention patent entitled "A method and device for scene feature extraction" on April 7, 2020 (application number: 202010265161.9), and the applicant was Beijing Xuanji Information Technology Co., Ltd.

在该专利中，提出了一种基于显著性评价的景物特征提取方法，比照人类对空间信息感知的过程，来提取目标景物的特征，以实现对景物中特定目标的准确识别。这里的显著性可以这样来理解：为了合理利用有限的视觉信息处理能力，人类需要选择整个视觉区域中特定的部分来集中关注，视觉中特定的部分称为显著性区域。

视觉的显著性区域在人类处理复杂的场景信息时起到了至关重要的作用。当面对一个复杂场景时，人类首先快速地浏览场景中的全部内容，利用全局的空间信息寻找场景中最重要的部分。然后集中注意力针对这一部分进行深度感知，实现对场景视觉结构的理解。

Based on the relevant information currently available on the plan, let us take a look at this technical solution.

As shown in the figure above, it is a flow chart of the scene feature extraction method invented in this patent. First, the system obtains the target image and performs binarization processing. Before binarization, the target image can also be filtered using, for example, mean filtering, median filtering, Gaussian filtering, etc. Secondly, the binary image is morphologically processed to obtain a morphologically processed image. Morphological processing includes: image corrosion operations, dilation operations, etc., which are used to eliminate small objects, separate objects at thin points, and smooth the boundaries of larger objects without significantly changing their areas.

Next, the excitation map is determined based on the morphologically processed image, that is, the excitation map of the target scene is obtained on the morphologically processed image using the Markov chain method. Specifically, for each point in the morphologically processed image, the gray value of the point is used as the state in the Markov chain, and the weight of the edge starting from the point is calculated. The weight of the edge is the probability of transition between states. By calculating the probability distribution of the Markov chain, the required excitation map is obtained.

最后，对激励图进行均一化处理得到显著性图，并根据显著性图确定目标图像的目标景物特征。

如上图，为该专利中展示的一种显著性图的示意图，其中，圆圈表示图中的亮色（白色，像素值为255）对应显著性区域，叉表示图中的暗色（黑色，像素值为0）对应图非显著性区域。根据显著性图和目标图像的映射关系，确定目标图像的显著性区域，进而针对目标图像的显著性区域进行后续识别处理，例如可以进一步识别该目标图像中包含的目标景物的类型、名称、状态等。

以上就是旋极信息发明的基于显著性评价的景物特征提取方案，该方案基于显著性评价提取的景物特征，能够更可靠地标识目标景物，可以适应图像局部灰度变化、尺度差异和微小形状变化，同时抗干扰性能也更强。