| 查看: 351 | 回复: 3 | |||
| 当前主题已经存档。 | |||
杨小军铁杆木虫 (著名写手)
|
[交流]
请求修改一段中译英 【计算机视觉 图像匹配】
|
||
|
为了提高图像拼接过程中的特征匹配准确率和自动化水平，提出了一种基于SIFT（Scale Invariant Feature Transform）特征的新匹配策略。首先，给定一个通用的匹配阈值，对SIFT特征对进行粗略判断，得到含有误匹配的特征对集合。其次，以特征对之间的欧式距离最小值与次小值之比为依据，取比率值最小的前8个特征对的图像坐标数据，求解图像透视变换参数初值，并计算这8个特征对的变换坐标与实际坐标之间的相对最大误差值σ。第三，计算上述8个特征对的原始图像坐标的最大分布范围，取其与图像尺寸的比值作为匹配误差门限控制参数k。最后，计算集合中的特征对变换坐标与实际坐标的差值，以该差值不大于3kσ作为控制条件，剔除误匹配，得到准确的匹配对集合，用于计算透视变换参数值。实验结果表明：在待匹配图像有一定程度的视点、光照、旋转、比例变化等情形下，该方法具有稳定、可靠的特点，所用实验图像的匹配准确率达到100%。该方法能无需人工选择匹配阈值，有效地提高了图像匹配的自动化水平。 In order to improve the matching accuracy and the level of automation for image mosaic, a new matching algorithm is proposed based on SIFT (Scale Invariant Feature Transform) features. Firstly, according to cursory comparing with the given common matching threshold, the corresponding SIFT characteristics collection which contains outlier is obtained. Secondly, taking into account the ratio of the minimum to second min Euclidean distance of corresponding features, get the image coordinates data of corresponding SIFT characteristics with first eight smallest ratios to solve the initial parameters value for image perspective transformation and then calculate the maximum relative error σ between the transformation coordinates data and the actual coordinates data of the eight corresponding characteristics. Thirdly, calculate the largest original image coordinates range of the above eight corresponding characteristics, the ratio to the image size is the matching error threshold k. Finally, compute the difference of the transformation coordinates data and the actual coordinates data of the corresponding characteristics in the characteristics collection, delete the corresponding characteristics whose difference are larger than 3kσ, then obtain the exact matching characteristics collection to solve the parameters value for image perspective transformation. The experimental results prove that the proposed method is stable and reliable under some view point, illumination, rotation and scale variation. This new method achieves 100% matching accuracy on the experimental images. Moreover, the proposed method can select the matching threshold of different images automatically and without manual intervention. [ Last edited by 杨小军 on 2010-3-2 at 13:17 ] |
回复此楼» 猜你喜欢
参与限项
已经有3人回复
-
假如你的研究生提出不合理要求
已经有7人回复
-
实验室接单子
已经有4人回复
-
全日制（定向）博士
已经有4人回复
-
对氯苯硼酸纯化
已经有3人回复
-
求助:我三月中下旬出站，青基依托单位怎么办？
已经有12人回复
-
不自信的我
已经有12人回复
-
所感
已经有4人回复
-
要不要辞职读博？
已经有7人回复
-
北核录用
已经有3人回复
杨小军
铁杆木虫 (著名写手)
- 应助: 27 (小学生)
- 金币: 8251.7
- 散金: 140
- 红花: 3
- 帖子: 1188
- 在线: 542.9小时
- 虫号: 844911
- 注册: 2009-09-10
- 性别: GG
- 专业: 系统科学与系统工程
2楼2010-03-02 13:19:03
杨小军(金币+10):非常感谢，改得很好 2010-03-03 12:15
杨小军(金币+10): 2010-03-04 12:07
杨小军(金币+10): 2010-03-04 12:07
| In order to improve the matching accuracy and the level of automation for image mosaic, a new matching algorithm is proposed based on SIFT (Scale Invariant Feature Transform) features. Firstly, according to cursory comparing with the given common matching threshold, the corresponding SIFT characteristics collection containing outlier is obtained. Secondly, considering the ratio of the minimum to second min Euclidean distance of corresponding features, the image coordinates data of corresponding SIFT characteristics with first eight smallest ratios are taken to solve the initial parameters value for image perspective transformation and then calculate the maximum relative error σ between the transformation coordinates data and the actual coordinates data of the eight corresponding characteristics. Thirdly, the largest original image coordinates range of the above eight corresponding characteristics is calculated and the ratio of which to the the image size is determined to be the matching error threshold k. Finally, the difference of the transformation coordinates data and the actual coordinates data of the corresponding characteristics in the characteristics collection is computed and the corresponding characteristics having greater difference than 3kσ are deleted. Then the exact matching characteristics collection is achieved to solve the parameters value for image perspective transformation. The experimental results indicate that the proposed method is stable and reliable in case of having viewpoints, illumination rotation and scale variation to some degree and this new method achieves 100% matching accuracy on the experimental images. Moreover, the proposed method can select the matching threshold of different images automatically with the absence of manual intervention. |
3楼2010-03-03 11:17:41
4楼2010-03-03 11:19:56