The present invention enables a high-quality filtered image to be generated even from multimodal and multispectral images containing positional deviations. An image perturbation part generates a perturbed guide image group comprising first to K-th perturbed guide images obtained by deforming a guide image. A filtering part generates a filtered image group comprising first to K-th filtered images by applying first to K-th filtering processing to a target image by using the perturbed guide image group. A reliability calculation part calculates a reliability group comprising first to K-th reliabilities for the first to K-th filtered images of the filtered image group on the basis of first to K-th correlation values between the first to K-th perturbed guide images and the target image. A weight optimization part generates, on the basis of the first to K-th reliabilities, a weight group comprising first to K-th weights to be respectively used when compositing the first to K-th filtered images. An output image compositing part composites an output image from the weight group and the filtered image group.