this change seems wrong to me.

we should keep

if (mat.empty())
    {
        mat.create( size.height, size.width, type );
    }
    else
    {
        CV_CheckEQ(size, mat.size(), "");
        CV_CheckTypeEQ(type, mat.type(), "");
        CV_Assert(mat.isContinuous());
    }

create should not reallocate data if requested image size fits into existing Mat.

Mat img(1000, 1000, CV_8UC3);
imread(f, img); // should not reallocate data if image size is less than (1000, 1000)

Without this change it will not be possible to read multiple images of different sizes efficiently, because user would need to pass either empty Mat or a Mat with exact image dimensions on each iteration.

I.e. with proposed change, following code should work and should perform reallocations only when necessary:

Mat img;
for (auto f : filenames)
{
    imread(f, img); // will fail if images have different sizes
}

sorry you are right. mat.create( size.height, size.width, type ); does the job.

    TickMeter tm;

    Mat im;

    for (int i = 0; i < 5; i++)
    {
        tm.reset();
        tm.start();
        imread("read.png", im);
        tm.stop();
        printf("read time %f\n", tm.getTimeSec());
    }
    imshow("read.png", im);
    waitKey();

read time 0.692092
read time 0.615938
read time 0.615163
read time 0.615883
read time 0.617650

good job!

InputOutputArray is better. I know, that they are synonyms, but it's simpler for reader.

It seems that it won't work without changing public imread to accept InputOutputArray (https://docs.opencv.org/4.x/d0/d46/classcv_1_1__InputOutputArray.html). There is no obvious way to construct InputOutputArray from OutputArray (https://docs.opencv.org/4.x/d2/d9e/classcv_1_1__OutputArray.html).