猪舍图像局部亮度调整方法

Method of local brightness adjusting of pigpen image

针对猪舍视频监控场景中常常遭遇局部亮度不均衡而引起后继图像分析困难问题,提出一种图像自适应局部亮度调整法(ALLA)。首先为避免物体色度的干扰,原图像转换为YCbCr模型后,只利用亮度Y分量图像,根据最大类间方差(Ostu)将Y图像分为明暗二区;其次针对过亮和过暗的局部区域采用正弦函数进行非线性反向调整灰度值;最后,为评价处理后图像质量,提取测试图和基准图中的猪只对象轮廓边缘像素成对梯度值,通过假设检验判断二者差的集合均值是否存在明显变化。选择了典型3种猪舍环境图像,一种光线柔和,图像亮度质量较为理想;另外两种夜晚圈栏灯光和白天阳光导致景物本色出现偏差,即在光照强度高的局部区域灰度值低,反之则高。试验采用ALLA处理后的测试图像,测试PSNR(峰值信噪比)值均在31～78之间,表明没有引起图像质量显著下降;采用我们设计的假设检验方法,表明在显著性水平(α=0.95)时,测试图较之于标准灰度化图像有显著改变,因此,有利于后继的猪只目标分割工作。

The pigpen scene in video frames often suffer from local disproportion luminance, which leads to inconvenience in subsequent images analysis. In this paper, an adaptive local lightness adjusting algorithm（ALLA） is proposed. Firstly,original RGB （red, green, blue） image is converted into YCbCr space （luminance is denoted by Y, Cb and Cr are the blue-difference and red-difference） in order to avoid the interference from chroma in YCbCr space. Secondly, only the Y gray-scale image of YCbCr space is divided into 2 areas of light and dark by adopting Ostu method. Thirdly, a method of nonlinear-reverse adjustment based on sine mode is applied to improving the gray value in the corresponding zones（ i.e. the excessive bright or dark ones）. Finally, for evaluating the validity of luminance improvement, a method of hypothesis testing is put forward, i.e. the processed image by ALLA is viewed as the testing one, and another processed image by standard graying is viewed as the reference one for the same original pigpen image; paired gradients of each pixel of the same pig’s edge in both them are computed; all paired gradient differences forms a set; the mean of the set as a index of the image quality is judged whether there is a significant change through hypothesis testing. Three types of typical pigpen images as testing samples are chosen in experiments. One of them is that the illumination is gentle to result in the luminance quality is satisfactory. Others are that the evening lighting and sunshine in the pigpen can cause the deviation on nature luminance, i.e. in the Y gray-scale image the low gray value in the zone vs. the high illumination intensity, otherwise, the high gray value in the zone. PSNRs of the testing ones after using ALLA are between 31 and 78, i.e. the quality level of the testing ones don’t decrease significantly. Furthermore, it is verified that the luminance level of the testing one are better than one of the reference one with the significance level α=0.95 based on our method of hypothesis testing. Meanwhile, another experiment shows the converged contour of pig in one other testing one by using same level-set method is more approximated to its actual contour than the reference. The results prove that ALLA is helpful for subsequent works on pig target segmentation.