Abstract: A method of image completion comprises: constructing the image repair model and constructing a plurality of conditional generative adversarial networks according to a plurality of object types; inputting the training image corresponding to the plurality of objective types such that the plurality of conditional generative adversarial networks respectively conduct a corruption feature training; inputting the image in need of repair and respectively conducting an image repair through the plurality of conditional generative adversarial networks to generate a plurality of repaired images; and judging a reasonable probability of the plurality of repaired images through a probability analyzer, choosing an accomplished image and outputting the accomplished image through an output interface.

Abstract: An indoor scene structural estimation system and an estimation method based on deep learning network are provided. An indoor scene structural estimation system based on deep learning network includes a 2D encoder, a 2D plane decoder, a 2D edge decoder, a 2D corner decoder, and a 3D encoder. The 2D encoder receives an input image and encodes the input image. The 2D plane decoder is connected to the 2D encoder, decodes the encoded input image, and generates a 2D plane segment layout image. The 2D plane decoder is connected to the 2D encoder, decodes the encoded input image, and generates a 2D plane segment layout image. The 2D corner decoder is connected to 2D encoder, decodes the encoded input image, and generates a 2D corner layout image.

Abstract: An indoor scene structural estimation system and an estimation method based on deep learning network are provided. An indoor scene structural estimation system based on deep learning network includes a 2D encoder, a 2D plane decoder, a 2D edge decoder, a 2D corner decoder, and a 3D encoder. The 2D encoder receives an input image and encodes the input image. The 2D plane decoder is connected to the 2D encoder, decodes the encoded input image, and generates a 2D plane segment layout image. The 2D plane decoder is connected to the 2D encoder, decodes the encoded input image, and generates a 2D plane segment layout image. The 2D corner decoder is connected to 2D encoder, decodes the encoded input image, and generates a 2D corner layout image.

Abstract: A method of image completion comprises: constructing the image repair model and constructing a plurality of conditional generative adversarial networks according to a plurality of object types; inputting the training image corresponding to the plurality of objective types such that the plurality of conditional generative adversarial networks respectively conduct a corruption feature training; inputting the image in need of repair and respectively conducting an image repair through the plurality of conditional generative adversarial networks to generate a plurality of repaired images; and judging a reasonable probability of the plurality of repaired images through a probability analyzer, choosing an accomplished image and outputting the accomplished image through an output interface.

Abstract: An image processing method applicable to images captured by a wide-angle zoomable lens is provided. Said method comprising steps of: A. utilizing the wide-angle zoomable lens to photograph a calibration pattern under a predetermined focal length to obtain a distorted image corresponding to the calibration pattern; B. calculating a distortion parameter for the wide-angle zoomable lens at the predetermined focal length; C. utilizing the wide-angle zoomable lens to photograph an object under another focal length to obtain another distorted image corresponding thereto; D. calculating a new distortion parameter by using the distortion parameter obtained from Step B and a multiple relation between said another focal length and the predetermined focal length; and E. using the new distortion parameter to perform distortion correction on said anther distorted image corresponding to the object. The present invention can perform distortion correction for real-time images captured by the wide-angle zoomable lens.

Abstract: A method and an image processing device for adjusting a stereo image are provided. The method includes: obtaining a disparity map between a first image and a second image; dividing the disparity map into layers which correspond to first layers of the first image and second layers of the second image; generating a first cost according to displacements of pixels of an ith first layer and an ith second layer; generating a second cost according to a difference between displacements of pixels of the ith first layer and a jth first layer, and a difference between displacements of pixels of the ith second layer and a jth second layer; executing an optimization algorithm at least according to the first cost and the second cost to obtain all the displacements; and adjusting the positions of the corresponding pixels according to the displacements. Therefore, a viewer may have a better viewing experience.

Abstract: A method to build a 3D model for a physical object includes the following steps. First, depth information of a plurality of images is converted into a plurality of 3D point clouds. Then, the motion parameters of the camera to gather each image relative to a previous image are determined according to the 3D point clouds. Next, a registration energy estimate value of each image's 3D point cloud, as being integrated into a previous image's 3D point cloud, is determined according to corresponding motion parameters. Then, the motion parameters are varied to minimize the estimate value. An optimal camera path is determined according to the varied motion parameters. Finally, a 3D model of the physical object is built according the optimal camera path.

Abstract: A method to build a 3D model for a physical object includes the following steps. First, depth information of a plurality of images is converted into a plurality of 3D point clouds. Then, the motion parameters of the camera to gather each image relative to a previous image are determined according to the 3D point clouds. Next, a registration energy estimate value of each image's 3D point cloud, as being integrated into a previous image's 3D point cloud, is determined according to corresponding motion parameters. Then, the motion parameters are varied to minimize the estimate value. An optimal camera path is determined according to the varied motion parameters. Finally, a 3D model of the physical object is built according the optimal camera path.

Abstract: A facial expression recognition system includes a facial database, an image capture unit, an expression recognition data generation unit and an expression recognition unit. The facial database includes a plurality of expression information and expression features of optical flow field, wherein each of the expression features of optical flow field corresponds to one of the expression information. The image capture unit captures a plurality of facial images. The expression recognition data generation unit is coupled to the image capture unit and the facial database for receiving a first facial image and a second facial image from the image capture unit and calculating an expression feature of optical flow field between the first facial image and the second facial image corresponding to each of the expression information.

Abstract: A method for correcting a rolling shutter effect is provided. The method includes: obtaining feature point pairs in images, wherein each of the feature point pairs corresponds to a motion vector; obtaining sampling points between two consecutive images in time; setting a moving velocity and an angular velocity of an image capturing unit at each of the sampling points as variables; obtaining estimating motion vectors according to the variables, a focal length of the image capturing unit, and row locations where the feature point pairs are located; executing an optimization algorithm according to a difference between the motion vectors and the estimating motion vectors, to calculate the moving velocity and the angular velocity corresponding to the variables; varying locations of pixels in an image according to the moving velocity and the angular velocity, to generate a first corrected image. Thereby, the rolling shutter effect in an image is removed.

Abstract: An image processing method applicable to images captured by a wide-angle zoomable lens is provided. Said method comprising steps of: A. utilizing the wide-angle zoomable lens to photograph a calibration pattern under a predetermined focal length to obtain a distorted image corresponding to the calibration pattern; B. calculating a distortion parameter for the wide-angle zoomable lens at the predetermined focal length; C. utilizing the wide-angle zoomable lens to photograph an object under another focal length to obtain another distorted image corresponding thereto; D. calculating a new distortion parameter by using the distortion parameter obtained from Step B and a multiple relation between said another focal length and the predetermined focal length; and E. using the new distortion parameter to perform distortion correction on said anther distorted image corresponding to the object. The present invention can perform distortion correction for real-time images captured by the wide-angle zoomable lens.

Abstract: A method and an image processing device for adjusting a stereo image are provided. The method includes: obtaining a disparity map between a first image and a second image; dividing the disparity map into layers which correspond to first layers of the first image and second layers of the second image; generating a first cost according to displacements of pixels of an ith first layer and an ith second layer; generating a second cost according to a difference between displacements of pixels of the ith first layer and a jth first layer, and a difference between displacements of pixels of the ith second layer and a jth second layer; executing an optimization algorithm at least according to the first cost and the second cost to obtain all the displacements; and adjusting the positions of the corresponding pixels according to the displacements. Therefore, a viewer may have a better viewing experience.

Abstract: A method for correcting a rolling shutter effect is provided. The method includes: obtaining feature point pairs in images, wherein each of the feature point pairs corresponds to a motion vector; obtaining sampling points between two consecutive images in time; setting a moving velocity and an angular velocity of an image capturing unit at each of the sampling points as variables; obtaining estimating motion vectors according to the variables, a focal length of the image capturing unit, and row locations where the feature point pairs are located; executing an optimization algorithm according to a difference between the motion vectors and the estimating motion vectors, to calculate the moving velocity and the angular velocity corresponding to the variables; varying locations of pixels in an image according to the moving velocity and the angular velocity, to generate a first corrected image. Thereby, the rolling shutter effect in an image is removed.

Abstract: A hierarchical motion deblurring method for a single image is provided. In the method, a blur kernel of a target image is calculated and a multi-scale representation for representing the target image and the blur kernel is constructed. Then, a gradient attenuation function and a strong edge suppression function are applied to a residual Richardson-Lucy algorithm, so as to iteratively calculate a residual image between the blur kernel and the target image represented by the representation in each scale and restore the residual image to obtain a first restored image and a second restored image. Finally, the two restored images are compared so as to obtain a motion deblurring image.

Abstract: A motion estimation method for video compression comprises the following steps. First, an initial simplex comprising three points is determined based on motion vectors in blocks of a current frame and a previous frame, and a point having a largest function value among the three points is replaced with a point having a smaller function value to form a simplex. The replacement is repeated until two points of the three points of the simplex converge to a same point. The iteration is performed by downhill simplex search including operations of reflection, expansion, contraction and shrinkage to find a point for replacement. The motion estimation method for video compression can also use multi-reference frames.

Abstract: A method for achieving segmentation of a picture according to one aspect of the present invention comprises: determining a first foreground of a picture based on a predetermined mask; applying Gaussian Mixture Models with weighted data (GMM-WD) to the first foreground to generate a second foreground; determining a first background of the picture based on the second foreground; applying the GMM-WD to the first background to generate a second background; and determining an unknown region based on the second background and the second foreground.

Abstract: The invention discloses a face recognition method that reconstructs a 3D face model from a single face image, synthesizes a set of face images under different conditions (such as pose, light . . . ) via the 3D face model, feeds the set of face images under different conditions to the face recognition classifier for training, and making intermediate decisions whether to-be identified individual from a series of video frames is a legal system user by the face recognition classifier. Moreover, the method not only recognizes legal system users, but also rejects imposters, a function inspired by the idea of LLE. Finally, better reliability can be achieved by fusing temporal intermediate decisions.

Abstract: The present invention provides a method for retargeting an image, comprising determining a total block structure energy of an input image content. A compressibility rate of the input image content is determined based on the total block structure energy. An optimal scaling factor of the input image content is obtained. The input image content is warped by using a new coordinate matrices and uniformly scaling the input image content to a target image resolution.

Abstract: A facial expression recognition system includes a facial database, an image capture unit, an expression recognition data generation unit and an expression recognition unit. The facial database includes a plurality of expression information and expression features of optical flow field, wherein each of the expression features of optical flow field corresponds to one of the expression information. The image capture unit captures a plurality of facial images. The expression recognition data generation unit is coupled to the image capture unit and the facial database for receiving a first facial image and a second facial image from the image capture unit and calculating an expression feature of optical flow field between the first facial image and the second facial image corresponding to each of the expression information.

Abstract: A hierarchical motion deblurring method for a single image is provided. In the method, a blur kernel of a target image is calculated and a multi-scale representation for representing the target image and the blur kernel is constructed. Then, a gradient attenuation function and a strong edge suppression function are applied to a residual Richardson-Lucy algorithm, so as to iteratively calculate a residual image between the blur kernel and the target image represented by the representation in each scale and restore the residual image to obtain a first restored image and a second restored image. Finally, the two restored images are compared so as to obtain a motion deblurring image.