How to remove 1.#QNAN and INF error in gradient (float type) in image in case of inf?

See more:

C++

OpenCV

image-processing

In a code to extract text from image (Stroke Width Transform), I am using gradients at specific pixels. In the following code: gradientX, gradientY are Mat images. G_x and G_y are gradients at (row, col) pixel. I am getting error because G_x and G_y are becoming out of bounds (1.#QNAN 1.INF) in many cases. In some cases, both become such that mag becomes INF or 1.#QNAN .

Please suggest some change in code such that this problem can be fixed. I would like to normalize them to 1 in case of INF rather than going out of bounds.

float G_x = gradientX.ptr<float>(row)[col];
        float G_y = gradientY.ptr<float>(row)[col];
        // normalize gradient
        float mag =  sqrt( (G_x * G_x) + (G_y * G_y) );
        if (dark_on_light){
            G_x = -G_x/mag;
            G_y = -G_y/mag;
        } else {
            G_x = G_x/mag;
            G_y = G_y/mag;

        }