This paper presents a digital three dimensional reconstruction method based on a set of small-baseline elemental images captured with a micro-lens array and a CCD sensor. object points relatively to the reference image frame. Using optimization algorithm with redundant matching points can achieve 3D reconstruction finally. Our experimental results are presented to demonstrate excellent performance in accuracy and speed of the proposed algorithm. lenslet. Each of the lenslet has a diameter of φ and 100% fill factor is assumed. The gap between the micro-lens array 3-deazaneplanocin A HCl and the camera sensor is . By using the triangular relationship the mapping between the object and image points is given by: Fig.1 Schematic layout of an integral imaging setup and the 3-deazaneplanocin A HCl geometric relationship between object point and image point is diameter of the lenslets and is the gap between the micro-lens array and the camera sensor. A 100% fill factor is assumed for the micro-lens array. For two different lenslets the equation (1) can be rewritten as and are sizes of sampling window. (+1)×(2× +1) lens array we calculate all the depth information with respect to the center reference image. The optimized depth value is obtained by computing a similarity criterion of by changing the value z each time: on the curve is regarded as the extracted depth of the object point. It is worth pointing out however the NCC method alone has high computation cost due to the iteration nature of the algorithm. To meet the need of real time applications in this paper we adopt ASIFT operator to improve computation efficiency which is further explained in Section 3. 3 ASIFT algorithm When the relative space position of two elemental images changes smaller light intensity is good enough SIFT(Scale-invariant feature transform) algorithm in image matching performs very well. But for 3D reconstruction match points we can Rabbit Polyclonal to DBF4. get are too sparse according to previous experiments. In our paper we use Figure 3 as system structure and experimental array image. The distance between the micro-lens and the 3D object are about 200mm (Cube) and 500mm (Kidney). The size of lens array is 100mm*100mm with each lens diameter at 1mm and the focal length is 3mm. In order to ensure the number of elemental image and resolution we chose 200*200 pixels as resolution of each image. Fig.3 System Structure and Images array 3-deazaneplanocin A HCl After cutting same size elemental images we use SIFT algorithm to extract match points [10-13]. As is shown in Figure 4 there are only 30 pairs of points with some error match points in bad light condition. Fig.4 Match points (including error match points) We cannot achieve 3D dense reconstruction by only 30 pairs of points including mismatch points. So we adopt ASIFT algorithm. But it does mean SIFT can be used instead of the ASIFT when a high-intensity light source. The main reason is as described in Figure 5 ASIFT simulates all distortions caused by the variation of the camera optical axis because the algorithm introduces two more parameters in order to achieve full affine invariant [14]. Then we can get more match points even in lower light condition. In the end we use the same method to match corresponding points as SIFT. In other words ASIFT simulates three parameters: scale camera longitude angle and latitude angle (which is equivalent to the tilt) and also normalizes the other three parameters (translation and rotation). This is affine invariant in the true sense [15]. Fig.5 Overview of the ASIFT algorithm [5]. The square images A and B represent the compared images u and v. ASIFT simulates all distortions caused by a variation of the camera optical axis direction. Affine Transformation Matrix A can be decomposed into : cos 1/ are the camera optical axis longitude and latitude respectively. The image is a flat physical object. The 3-deazaneplanocin A HCl small parallelogram on the top represents a camera looking at . A third angleψ parameterizes the camera spin and λ is a zoom parameter. Fig.6 Camera model In order to have ASIFT invariant to any affine transformations we need to sample the tilt and angle Φ with a high enough precision. The sampling steps Δand ΔΦ must be fixed experimentally by testing several natural images. Figure 7 illustrates the irregular sampling results: θ and Φ on the observation hemisphere where ((on the axis. For digital images tilt images are determined by directional t-subsampling. It requires antialiasing filter processing on the axis in order to minimize the distortion of the image. The filter is performed by Gaussian convolution which standard deviation is = 0.8 . In the reference [5] it proved.