基于尺度不变特征转换算子的水果表面图像拼接方法

Matching method for fruit surface image based on scale invariant feature transform algorithm

水果全表面图像信息是否完整,直接影响水果表面颜色和缺陷检测的结果。该文提出了一种基于尺度不变特征转换(SIFT,scale invariant feature transform)算子的图像拼接方法,实现多视角水果图像的拼接以获取完整的水果表面信息。首先以15°固定间隔旋转水果以获取各视角下的连续图像,在图像2R-G-B通道下实现图像目标和背景分离,并对目标图像进行灰度直方图均衡化以增强其纹理信息,有利于特征点的提取。运用SIFT算法提取图像特征点,因为特征向量数量多、维数高,采用普通的K-D树算法搜索匹配点将消耗大量时间,因此将图像划分为16个区域,通过多次试验可知中间4个区域为特征点是最容易匹配的区域,这样就缩小匹配点可能存在的区域。采用极线几何约束法和改进型随机抽样一致(random sample consensus,RANSAC)算法以提高图像拼接精度,减少匹配时间。根据平移矩阵,对前后图像进行拼接,从而实现水果表面图像的完整拼接。试验结果表明:该算法平均匹配精度提高35.0%,平均拼接时间为2.5 s,较传统K-D树算法缩短67.8%时间,拼接效果还原率为93.9%。该文算法具有一定的尺度、旋转以及仿射变换不变性,适用于随机呈现的不同姿态球状水果图像拼接。该研究可为基于机器视觉的农产品品质检测和等级划分提供科学参考。

The completely fruit surface image information is an important factor which will directly influence the detection results of fruit＇s surface color and defect. This paper took the common red delicious apple as the research object. An image feature extraction and matching method based on SIFT algorithm was proposed, and the multi-view fruit image were stitched effectively in this paper. The algorithm was helpful to obtain the completely fruit surface image information. Firstly, the fruits were rotated at fixed interval 15° angle and the multi-view of fruit continuous images was achieved. Based on the analysis of fruit image color space, the fruits target and the background were divided by 2 R-G-B channels for removing image noise. The target image was proposed by gray histogram equalization; hence the image＇s contrast was enhanced. The pre-paired image had special information which could be used for extracting feature points. After comparing with speeded-up robust features （SURF） and scale invariant feature transform （SIFT） algorithm, image feature points were detected between two images using SIFT algorithm. The average number of characteristic vector with 128 dimensions for each image was 2500. Because of large quantity and high dimensions of characteristic vector, significant amount of time was consumed when using the traditionalK-D tree algorithm in searching matching points. To reduce the matching point of the existing area, a complete fruit image was divided into 16 regions, and four regions in the middle area with the most easily matching area for feature points were selected by multiple tests. A series of images collected by CCD camera only had lateral deviation between pre and post image. The searching scope of matching point was controlled in a narrow space between ±10 pixels through epipolar geometric constraint algorithm. Therefore, the mismatching rate was reduced and the images matching precision was improved. Finally, the mismatching points were rejected using the improved random sample consensus （RANSAC） algorithm, and it also could be further improved for matching precision. The initial translation matrix was obtained through rough matching points. The euclidean distance between pre and post image matching points selected randomly by using RANSAC was calculated. It was helpful to distinguish the interior point and exterior point. The final precisely matching points for each image were obtained based on presupposition threshold condition and point number condition. Calculating the center coordinates of the final match points for each image, the image of around the center coordinates for pre and post treated were reserved, and the translation matrix was generated at the same time. According to the translation matrix, the complete fruit surface image stitching was realized though stitching the images characteristics of pre and post. The experimental results indicated that the matching algorithm could dramatically reduce the mismatching rate and improved the average matching precision by 35.0%, the average matching time decreased from 7.8 s to 2.5 s, and the reduction rate was 67.8% compared with traditionalK-D tree algorithm, and the reduction rate of matching results was 93.9%. This algorithm was also effective for the arbitrary pose fruit image on the test bench. This algorithm had good real-time performance, and it was invariance to scale, rotation, and affine transform, and it was effective for the randomly pose of spherical fruit images matching. This study provides an important reference for the quality detection and grade division of agricultural products base on machine vision.