无人机模糊图像自动检测方法

图像模糊程度是图像评价的一个关键指标参数,它影响着图像的质量和信息。特别对于无人机拍摄的图像,如果利用模糊的图像参与计算,会造成很大的误差甚至出现颠覆性的结果。因此对于图像模糊的检测显得至关重要。传统的图像模糊检测方法大都基于人工检测、且有合格的参考图像参与评价过程,方法耗时、费力,无法用于大量无人机拍摄图像的分析。因此本文基于Sobel边缘检测原理,利用4个方向的Sobel算子,寻找图像中每个Sobel边缘点的模糊邻域,并构建模糊邻域宽度值的计算准则,由此来计算出整幅图像的平均模糊邻域宽度值,并将这个计算结果作为检测模糊的直接依据。同时考虑到无人机拍摄图像的特点,将其按照拍摄时间顺序排列,依次将相邻图像互为参考,通过对比互为参考图像的模糊邻域宽度值的变化情况,将宽度值突变的图像确定为模糊图像。据此对无人机拍摄的所有图像进行模糊检测。最终通过7组2322张 图像进行自动检测发现151张图像模糊,通过人工检测发现158张图像模糊,平均检测率95.57%。该检测方法具有较强的 适用性。

The Research on Automatic Blur Detection in UAV Image

With the development of technology, the Unmanned Aerial Vehicle (UAV) is widely used for scientific activities. Image acquisition is its main function. The bad image quality can affect processing results. Image blur is a key indicator of image evaluation. It determines quality and accuracy of the image. If the blurred UAV images are used in the subsequent calculation and processing, the results will be unreliable and unrobustness. This is even a serious error. As a result, the detection of blurred image is of great significance to use. The reason of image blurring has 4 kinds of factors. They are weather conditions, UAV platforms, camera systems and environments. The weather conditions mainly include rain and wind. The UAV platforms mainly are GPS signal intensity. They can make unstable position for UAV. The camera systems mainly include parameters setting of camera, such as focal length, ISO, shutter speed and aperture value. Environments mainly include terrain and illumination. Undulating terrain and different illumination intensity maybe lead to focus inaccuracy. These factors also make image blurring. According to the cause of blurring, there are motion blur and defocus blur. The traditional methods of detecting image blur are mainly based on manual inspection and qualified reference images in the process of evaluation. However, this method is considerably time-consuming and laborious. It is not suitable for a great number of UAV images processing. This paper used the four directions of Sobel edge detection algorithm for building basic evaluation principle, finding blur neighborhood width of each Sobel edge detection point in whole image. Finally, constructing the calculation guidelines of blur neighborhood width. The average value of the blur neighborhood width is calculated by sum and average operation for each sobel edge detection point, and used this value as a direct basis of detecting image blur. The average blur neighborhood width is a dimensionless value. It is affected by richness of image information. So it cannot be used to direct comparison as an absolute reference value. It can also be used for relatively comparison between the similar images, which has approximate richness of image information. Meanwhile, taking into account the characteristics of the UAV images, they have a certain overlap and series. We put time-adjacent images as mutual referenced images because the time-adjacent images have the similar richness of image information. By relatively comparing the changes of blur neighborhood width, when the change is more than a certain threshold, the blurred images have been determined. According to this method, the whole image has been detected. Through a number of experiments, the whole 4 thresholds have been determined. There are m=5, T=5, T₁=0.2 and T₂=0.167. These thresholds can also be used to UAV image blur detection. Finally, we processed 2322 images of different feature types. There are hill, urban, mountain and plain. They were divided into 7 groups with automatic detection. 151 images was blurred while 158 images was blurred by manual inspection as the correct detection results. The average rate of detection was 95.57%. The detection method has a strong applicability.