Exception with OpenCv DnnSuperResImpl (C++)

Question

I have a problem with the upsample function from the dnnsuperres pack of OpenCV. I am transferring a bmp via a pipe from one application to another. The image resides in datamem, which is an unsigned char array. I use the pipe to transfer many other things, so I transfer the data in the most common datatype. If I do the following:

        cv::Mat matImg;
        matImg = cv::imdecode(cv::Mat(1, SizeToRead, CV_8UC1, datamem), -1); //-1==CV_LOAD_IMAGE_UNCHANGED 
        imshow("SomeWindowName", matImg);

This succeeds without a problem, the bitmap is shown correctly.

If I now try to apply an upsampling:

        string path = "EDSR_x4.pb";
        string modelName = "edsr";
        int scale = 4;
        DnnSuperResImpl sr;
        sr.readModel(path);
        sr.setModel(modelName, scale);
        Mat outputImage;
        sr.upsample(matImg, outputImage);

sr.upsampling() will thrown an exception. Unfortunately the exception info doesn't help me any further (cv::exception at memory location xyz, no additional info. The address seems reasonable). So I tried to simply load a bitmap from my hdd via cv::imread() and pass this to sr.upsample() and this works. I think it might be some format of the bitmap in datamem which is incompatible with sr.upsampling(). Therefore I tried to convert the bitmap to a png and pass it to the upsampling method, but the result was the same, I got the exception. The conversion I did like this:

        matImg = cv::imdecode(cv::Mat(1, SizeToRead, CV_8UC1, datamem), -1); //-1==CV_LOAD_IMAGE_UNCHANGED
        cv::Mat test(matImg.rows,matImg.cols,CV_8UC4, (cv::Vec4b*)matImg.data);//convert to png

Afterwards I passed of course test to upsampling(), but I got the same exception.

Does anyone have an idea what is causing the problem?

Answer 1

the DnnSuperResImpl will only accept 3 channel, BGR images, not your 4 channel bitmaps.

convert your image to 3 channels, while decoding it:

matImg = cv::imdecode(cv::Mat(1, SizeToRead, CV_8UC1, datamem), cv::IMREAD_COLOR);
sr.upsample(matImg, outputImage);

So far I found no documentation which says that it needs three channels.

indeed, please raise an issue here