Adrian Zamfir

Low Complexity Global Motion Estimation Techniques for Image Stabilization

Two new global motion estimation methods are proposed. The first one, called sign projection (SP) is obtained by modifying the integral projection estimation method using two thresholds for pixel values. The second method, called binary incrementation (BI), is obtained by using only one threshold for generating binary vectors from two images. It is shown that the proposed approaches provide similar motion estimation accuracy with the integral projection (IP) and phase correlation (PC) methods. Also, they have reduced numerical complexity and memory requirements, leading to shorter processing times as well as lower power consumption. The technique is particularly suitable for implementation in consumer devices such as digital video cameras.

A Study of the Influence of the PSF Accuracy on the Quality of Image Deblurring

The paper presents a few recent results obtained in restoration of motion blurred photographs. Qualitative differences between images restored with point spread functions of different shapes, values or sizes are illustrated, using test images and simulated blur.

Adaptive recovery of motion blur point spread function from differently exposed images

Motion due to digital camera movement during the image capture process is a major factor that degrades the quality of images and many methods for camera motion removal have been developed. Central to all techniques is the correct recovery of what is known as the Point Spread Function (PSF). A very popular technique to estimate the PSF relies on using a pair of gyroscopic sensors to measure the hand motion. However, the errors caused either by the loss of the translational component of the movement or due to the lack of precision in gyro-sensors measurements impede the achievement of a good quality restored image. In order to compensate for this, we propose a method that begins with an estimation of the PSF obtained from 2 gyro sensors and uses a pair of under-exposed image together with the blurred image to adaptively improve it. The luminance of the under-exposed image is equalized with that of the blurred image. An initial estimation of the PSF is generated from the output signal of 2 gyro sensors. The PSF coefficients are updated using 2D-Least Mean Square (LMS) algorithms with a coarse-to-fine approach on a grid of points selected from both images. This refined PSF is used to process the blurred image using known deblurring methods. Our results show that the proposed method leads to superior PSF support and coefficient estimation. Also the quality of the restored image is improved compared to 2 gyro only approach or to blind image de-convolution results.

MPEG-7 compliant generalized structure descriptor for still image indexing

This paper proposes a generalized structure descriptor (GSD), that offers an MPEG-7 compliant template for the introduction of a large class of local image content descriptors. The standard MPEG-7 color structure descriptor (CSD) is a particular case of the proposed GSD. The GSD descriptor embed color and spatial distribution information, allowing a compact description of color and texture to be used in content-based image retrieval. We will present several particular cases of spatial image structure description. Experiments show that the proposed descriptors can be successfully used for color texture and general image recognition and retrieval.

Correction of image artifacts produced by airborne flash reflections

This paper presents a method for correcting digital image artifacts that appear in images acquired by digital cameras caused by flash light reflection on out of focus airborne particles. This method makes use of polar coordinates representation of artifacts and it is assumed that they were previously detected.

People first: separating people from background in digital photographs

Extracting people from background in digital photography is a task of great importance, with many applications for digital cameras. Yet, the task poses a number of challenging technical problems to be tackled. In this paper we propose a novel technique for people extraction from background which is both accurate and of low computational complexity therefore amenable to be embedded in digital cameras. The proposed technique uses frames from the camera live view mode (called previews) (now widely available in digital cameras and even in the latest DSLRs) in conjunction with the flash. The basic principle of the method is to acquire two images of the same scene, one with flash and the other without. The use of preview images over two captured images makes the solution easily embeddable in digital cameras. In the proposed setup, in daylight conditions, the flash is triggered at the time of the penultimate preview image. The mask of the subject is then computed based on the intensity difference between the last two previews. For night scenes, where the flash power is required for the acquisition of the actual picture, the subject is detected based on the intensity difference between the final image downsampled to the size of the preview and the average of the last two previews. Additional problems posed by this setup, e.g. misalignments, false positives, incomplete subject map, are also addressed. The resulting foreground map is further used to obtain a narrow depth-of-field version of the initial photograph, by keeping the foreground unaltered while blurring the background.