基于纹理特征的SIFT算法改进

Improved SIFT Algorithm Based on Texture Features

针对SIFT(Scale Invariant Feature Transform)算法中使用固定对比度阈值提出了改进方法。当红外图像纹理特征不明显时,算法所能提取的特征点数量会大量减少,影响后续利用特征点进行如图像匹配、目标识别等处理。而人工改变对比度阈值具有局限性,不适用于很多场合。因此提出了一种基于纹理特征的自适应对比度阈值的SIFT算法。所使用的纹理特征提取方法是灰度共生矩阵,鉴于灰度共生矩阵并不能直接应用的特点,因此提取了特征参数。在图像纹理的特征参数如角二阶矩较大时,调低对比度阈值,使得特征点数量得以提高。此算法经验证表明能够在图像纹理特征不明显的情况下依然提取出大量的SIFT特征点。

An improved method used for changing the fixed threshold in Scale Invariant Feature Transform algorithm is proposed. If the texture features of the infrared image is not obvious, the feature points will be significantly reduced, so it will influence the subsequent procedure such as image registration, object recognition and etc. Modifying the contrast threshold artificially is not adapt to many occasions because of its limitation. Therefore, an adaptive contrast threshold SIFT algorithm based on texture features is necessary. Gray level co-occurrence matrix is one of the methods to represent the texture features. Owing to its feature that it could not analyze the image directly, the characteristic parameter must be extracted. The contrast threshold is modified lower only when the characteristic parameter of texture such as angular second moment becomes larger, so that the characters can be much more. The results indicate that even when the texture features of an image is not clear, large numbers of Scale Invariant Feature Transform characteristics can still be extracted.