Find image within image (object detection)

 

I do have different images which all have some kind of border around the "real" image. What I would like to achieve is to find the "real" image (size and location in pixels).

真正的图像显示与鲨鱼的水

For me the challenge is that the border is not always black (can be any kind of black or grey with a lot of noise) and the "real" image (water with shark in this example) can have any combination of color, saturation, ...

Now in general I'm aware of algorithms like Canny, Blob detection, hough lines, ..., but I have just started using them. So far I managed to find the border for a specific image, but as soon as I try to apply the same algorithms and parameters to the next image it doesn't work. My current approach looks like this (pseudo code):

  • convert to gray CvInvoke.CvtColor(_processedImage, tempMat, CvEnum.ColorConversion.Rgb2Gray)
  • downsample with CvInvoke.PyrDown(srcImage, targetImage) and CvInvoke.PyrUp(srcImage, targetImage)
  • blur image with CvInvoke.GaussianBlur(_processedImage, bluredImage, New Drawing.Size(5, 5), 0)
  • Binarize with CvInvoke.Threshold(_processedImage, blackWhiteImage, _parameters.BinarizeThreshold, 255, CvEnum.ThresholdType.Binary)
  • Detect Edges with CvInvoke.Canny(_processedImage, imgEdges, 60, 100)
  • Find Contours with `CvInvoke.FindContours(_processedImage, contours, Nothing, CvEnum.RetrType.External, CvEnum.ChainApproxMethod.ChainApproxSimple)
  • Assume that largest contour is the real image

    • I already tried different approaches based on for example:

  • Thresholding saturation channel and bounding box
  • Thresholding, canny edge and finding contours

    • Any hint especially on how to find proper parameters (that apply for all images) for algorithms like (adaptive) threshold and canny as well as ideas for improving the processing pipeline would be highly appreciated.

    您可以尝试从该图像中减去黑色图像，并且您将获得内部图像，方法如下：使用图像相减来比较C＃中的图像，

    If the border was uniform, this would be easy. Use cv::reduce to find MIN and MAX of each row and column; then count the top,left,bottom,right rows/columns whose MIN and MAX are equal (or very close) to the pixel value in a nearby corner. For sanity, maybe check the border colour is the same on all sides.

    In your example the border contains faint red stuff, but a row/column approach might still be a useful way to simplify the problem. Maybe, as Nofar suggests, take an absolute difference with what you think is the background colour; square it, convert to grey, then reduce to Sums of rows and columns. You still need to find edges, but have reduced the data from two dimensions to one.

    If there's a large border and lots of noise, maybe iterate: in the second pass, exclude the rows you think comprise the border, from statistics on columns (and vice versa).

    EDIT: The above only works for an upright rectangle! If it could be rotated then the row/column projection method won't work. In that case I might go for sum-of-squared differences as above (don't start by converting to grey as it could throw away information), followed by blurring or some morphology, edge detection then some kind of Hough transform to find straight edges.

    链接地址: http://www.djcxy.com/p/13978.html

    上一篇: ASP.NET MVC应用程序的开源许可选项？

    下一篇: 在图像中查找图像（对象检测）