A common way to detect blur is to measure the loss of high-frequency detail. Sharp images contain strong local intensity changes, so first- and second-derivative operators are often used as blur metrics.

A practical approach is to filter the image with a gradient operator or, more commonly, a Laplacian (second derivative). If the response is weak overall, the image is likely blurred because fine detail and edges have been smoothed away. You can then summarize that response with a statistic such as average magnitude or variance to produce a blur score.

To specifically detect motion blur, look for directionality. Linear motion blur tends to suppress detail along one axis more than the other, so anisotropic filters can help reveal that pattern. Methods that analyze directional structure, such as corner/structure-based detectors like the Harris corner detector, can help distinguish motion blur from general softness because motion blur often leaves detail preserved in one direction while reducing it in another.

These methods are well suited to convolution-based implementation in C or MATLAB (for example, using 2D convolution functions such as conv2 in MATLAB).