If there are low-level noise with uniform distribution in the image, the image contains high-frequency variation which is not obvious but which will affect compression performance. After applying DCT, this variation produces a number of high-frequency AC coefficient, some of which remain after quantisation. This means that more bits will remain after compressing the noised image than compressing the clean image. For example, the original image compress to 3211 bytes and noised image compress to 4063 bytes. The compression efficiency reduces by 25%.
We can use pre-processing (pre-filtering here) the image data before encoding to improve compression efficiency. The aim of a pre-filter is to increase compression performance without adversely affecting image quality. For example, using a Gaussian 2D filter to the above noised image.
Camera shake (jitter) is another cause of poor compression efficiency. Block-based motion estimation performs best when the camera is fixed in one position or when it undergoes smooth linear movement. If the motion search algorithm does not detect the jitter correctly, the result is a large residual frame after motion compensation.
The compression efficiency may be increased with automatic camera stabilization. Either mechanical stabilization or electronic image stabilization (remove vibration)