Abstract: A method is introduced for detecting approximately-circular object in infrared image by using the multi-feature joint probability. Bending direction of short-length curves in the image is calculated to get the directing probability of circles existing at each pixel. Edge intensity in gradient image is used to weight the probability, which is then combined with the probability of hole existing at each position and get a final judge whether or where the approximate-circular objects occur. The method shows some advantages when compared with the resolution from Hough transformation. Experiments with some quite different infrared images show that the method can successfully detect approximately-circular objects and shows good stability, so it is useful for target detection and recognition in some application.