The present disclosure relates to a method, device, computer readable media, and electronic devices for identifying a traffic light signal from an image. The method for identifying a traffic light signal from an image includes extracting, based on a deep neural network, multiple layers of first feature maps corresponding to different layers of the deep neural network from the image. The method includes selecting at least two layers of the first feature maps having different scales from the multiple layers of the first feature maps. The method includes inputting the at least two layers of the first feature maps to a convolution layer having a convolution kernel matching a shape of a traffic light to obtain a second feature map. The method includes obtaining a detection result of the traffic light signal based on the second feature map.

Pixels are classified within a time series of first and second images for the first image, a first probability map is provided with a first probability for a cloud for each first pixel and, for the second image, a second probability map with a second probability for a cloud for each second pixel; first and second mean intensity values are calculated for the pixels; local zero mean images are calculated by subtracting the mean intensity value from the intensity value of the respective pixel; a maximum difference map is generated by calculating, for spatially corresponding pixels, an absolute difference value between a first and second zero mean value; a weighting map is produced by multiplying each absolute difference value with a non-linear function; and a classifying map is computed based on the first probability map, the second probability map, and the weighting map.

Disclosed is a fusion network-based method for image super-resolution and non-uniform motion deblurring. The method achieves, for the first time, restoration of a low-resolution non-uniform motion-blurred image based on a deep neural network. The network uses two branch modules to respectively extract features for image super-resolution and non-uniform motion deblurring, and achieves, by means of a feature fusion module that is trainable, adaptive fusion of outputs of the two branch modules for extracting features. Finally, an upsampling reconstruction module achieves a non-uniform motion deblurring and super-resolution task. According to the method, a self-generated set of training data is configured to perform offline training on a network, thereby achieving restoration of the low-resolution non-uniform motion-blurred image.

A method for processing images includes: detecting a plurality of human face key points of a three-dimensional human face in a target image; acquiring a virtual makeup image, wherein the virtual makeup image includes a plurality of reference key points, the reference key points indicating human face key points of a two-dimensional human face; and acquiring a target image fused with the virtual makeup image by fusing the virtual makeup image and the target image with each of the reference key points in the virtual makeup image aligned with a corresponding human face key point.

A method for detecting and locating a lesion in a medical image is provided. A target medical image of a lesion is obtained and input into a deep learning model to obtain a target sequence. A first feature map output from the last convolution layer in the deep learning model is extracted. A weight value of each network unit corresponding to each preset lesion type in a fully connected layer is extracted. For each preset lesion type, a fusion feature map is calculated according to the first feature map and the corresponding weight value and resampled to the size of the target medical image to generate a generic activation map. The maximum connected area in each generic activation map is determined, and a mark border surrounding the maximum connected area is created. A mark border corresponding to each preset lesion type is added to the target medical image.

A method for detecting boxes includes receiving a plurality of image frame pairs for an area of interest including at least one target box. Each image frame pair includes a monocular image frame and a respective depth image frame. For each image frame pair, the method includes determining corners for a rectangle associated with the at least one target box within the respective monocular image frame. Based on the determined corners, the method includes the following: performing edge detection and determining faces within the respective monocular image frame; and extracting planes corresponding to the at least one target box from the respective depth image frame. The method includes matching the determined faces to the extracted planes and generating a box estimation based on the determined corners, the performed edge detection, and the matched faces of the at least one target box.

The present invention extends to methods, systems, and computer program products for classifying time series image data. Aspects of the invention include encoding motion information from video frames in an eccentricity map. An eccentricity map is essentially a static image that aggregates apparent motion of objects, surfaces, and edges, from a plurality of video frames. In general, eccentricity reflects how different a data point is from the past readings of the same set of variables. Neural networks can be trained to detect and classify actions in videos from eccentricity maps. Eccentricity maps can be provided to a neural network as input. Output from the neural network can indicate if detected motion in a video is or is not classified as an action, such as, for example, a hand gesture.

The technology described in this document can be embodied in a method for generating an iris identifier. The method includes obtaining a plurality of images of an iris, and generating a binary code for each of the plurality of images of the iris, the binary code including a sequence of bits. The method also includes identifying a first pattern of bits for which bit values and bit-locations are the same across a plurality of the binary codes, generating a first index based on the first pattern of bits, and then storing the first index on a storage device in accordance with a database management system. The first index is linked to biometric information of a different modality for a corresponding user.

Aspects of the present disclosure involve a system and a method for performing operations comprising: receiving a two-dimensional continuous surface representation of a three-dimensional object, the continuous surface comprising a plurality of landmark locations; determining a first set of soft membership functions based on a relative location of points in the two-dimensional continuous surface representation and the landmark locations; receiving a two-dimensional input image, the input image comprising an image of the object; extracting a plurality of features from the input image using a feature recognition model; generating an encoded feature representation of the extracted features using the first set of soft membership functions; generating a dense feature representation of the extracted features from the encoded representation using a second set of soft membership functions; and processing the second set of soft membership functions and dense feature representation using a neural image decoder model to generate an output image.

Systems and methods are described for generating a three dimensional (3D) effect from a two dimensional (2D) image. The methods may include generating a depth map based on a 2D image, identifying a camera path, generating one or more extremal views based on the 2D image and the camera path, generating a global point cloud by inpainting occlusion gaps in the one or more extremal views, generating one or more intermediate views based on the global point cloud and the camera path, and combining the one or more extremal views and the one or more intermediate views to produce a 3D motion effect.