Realtek Semiconductor Inc.

Shih-Tse Chen

Tsung-Hsuan Li

Chao-Hsun Yang

Chien-Hao Chen

Hao-Tien Chiang

Chun-Chang Wu

Wen-Tsung Huang

Wan-Ju Tang

Shang-Lun Chan

Yang-Ting Chou

Chih-Wei Wu

Wei-Te Chang

Kai Yin Liu

Kai-Chun Su

Chih-Yu Hsu

Wei-Hsiang Shen

Shian-shin Lu

Min-Xuan Qiu

Ying-Liang Chou

Yu-Hsuan Kuan

A neural network system and a signal processing method are provided. The neural network system includes at least one processing unit and a neural network module. The signal processing method includes: inputting a neural network input to the neural network module by the processing unit to generate an input at a previous layer of each convolutional transformer layer; performing pointwise convolution on the input by a key embedding layer based on key convolutional kernels to output a key tensor; performing convolution on the input by a value embedding layer based on value convolutional kernels to output a value tensor; performing a convolution on the cascading tensor of a first tensor and the key tensor by an attention embedding layer based on attention convolution kernels to output an attention tensor; and outputting an output tensor based on the attention tensor and the value tensor by an output module.

Neural network system and signal processing method

A training system, a training method, and a recognition system are provided. The training method is used to train a neural network module including: an encoder module, a shared decoder module, a synthesis module, and a classification module. The training method includes performing in a training epoch: repeatedly executing: taking a training image from a training set as an input image, obtaining a first loss based on training feature images of the training image and the feature images corresponding to the training image, and obtaining a second loss based on a classification marker of the training image and a classification generated by the classification module in correspondence with the training image; and updating first parameters and second parameters based on an average value of all the first losses and an average value of all the second losses obtained in the preceding step and an update algorithm.

Training system, training method and recognition system

A method of data augmentation is provided. The method includes the following operations: selecting an original image from an original dataset including label data configured to indicate a labeled area of the original image; selecting at least part of content, located in the labeled area, of the original image as a first target image; generating a first sample image according to the first target image, in which the first sample image includes the first target image and a first border pattern different from the first target image, and the content, located in the labeled area, of the original image is free from including at least part of the first border pattern; and incorporating the first sample image into a sample dataset, in which the sample dataset is configured to be inputted to a machine learning model.

Method of data augmentation and non-transitory computer readable media

A method for determining deblur filter coefficients includes the following steps: generating an edge profile according to the data of a blurred image; estimating a blur kernel according to the edge profile, wherein the blur kernel indicates how an imaging process blurs original image data and thereby generates blurred image data; and determining coefficients of a deblur filter to make a process result of the deblur filter processing the blurred image data approach the original image data.

Method for determining coefficients of deblur filter

A detection system and a detection method are provided. The detection method includes: obtaining first, second and third keypoints of a face by a keypoint acquisition module based on an image containing the face, the keypoint acquisition module obtaining the third keypoint based on a predetermined position on a midline of a human face and the first and the second keypoints based on two paired positions outside the midline; obtaining a vector by a first calculation module based on the first and second keypoints; obtaining a two-variable linear function by a second calculation module based on the vector and the third keypoint; and substituting, by a determination module, the coordinates of the first keypoint and the coordinates of the second keypoint into the two-variable linear function to obtain first and second values, respectively, and determining the state of the face based on the first and second values.

Detection system and detection method

An object detection model training apparatus, method and non-transitory computer readable storage medium thereof are provided. The apparatus performs a first object detection on a plurality of training images to generate a piece of first label information corresponding to each of the training images by a first teacher model. The apparatus trains a student model based on the training images and the first label information. The apparatus performs a second object detection on the training images to generate a piece of second label information corresponding to each of the training images by a second teacher model. The apparatus trains the student model based on the training images and the second label information.

Object detection model training apparatus, method, and non-transitory computer readable storage medium thereof

The present invention discloses an image data augmentation method that includes the steps outlined below. At least one distortion operation function is retrieved. A plurality of pixels included in the image are twisted according to the distortion operation function to generate at least one augmented image. Object information of each of at least one object included in the image is converted according to the distortion operation function to generate object information conversion result. The augmented image, a class tag of each of the at least one object and the object information conversion result are fed to a machine learning module to generate a machine learning result.

Image data augmentation apparatus and method

An image data augmentation device includes a memory and a processor. The memory is configured to store multiple commands and multiple images; the processor is connected to the memory to read the multiple images and execute multiple commands to: identify at least one object-related image including at least one object from the multiple images; capture at least one object image from the at least one object related image; and superimpose the at least one object image on multiple arbitrary positions in the multiple images to generate multiple training sample images, and using the multiple training sample images to execute machine learning. In addition, an image data augmentation method is also disclosed here.

Image data augmentation device and method

An image processing method includes: receiving an input image; performing a low-frequency image regulating operation to regulate the local intensity of the image of pixel unit(s) according to low-frequency information of the image of pixel unit(s) of the input image; performing a high-frequency image regulating operation to improve the details of the image of pixel unit(s) according to high-frequency information of the image of pixel unit (s) of the input image; and, generating an output image according to the input image, the low-frequency image regulating operation, and the high-frequency image regulating operation.

Image processing apparatus and image processing method for contrast enhancement

A method used for object tracking includes: using a specific object model to generate a first vector of a first ratio object and a second vector of a second ratio object of an image in an object detection bounding box of a specific frame; generating an identity label of an object within the bounding box according to the first vector, the second vector, and M first ratio reference vectors and M second ratio reference vectors stored in an object vector database.

Electronic apparatus and method used for object tracking

A method of performing person re-identification includes: obtaining a person feature vector according to an extracted image containing a person; obtaining state information of the person according to a state of the person in the extracted image; comparing the person feature vector with a plurality of registered person feature vectors in a database; when the person feature vector successfully matches a first registered person feature vector of the plurality of registered person feature vectors, identifying the person as a first identity corresponding to the first registered person feature vector; and selectively utilizing the person feature vector to update one of the first registered person feature vector and at least one second registered person feature vector that correspond to the first identity according to the state information.

Method and apparatus for person re-identification

A character recognition method includes the following operations: determining that the image of character to be identified corresponds to a matching character of several registered characters according to several vector distances to be identified between a vector of an image of character to be identified and several vectors of several registered character images of several registered characters, and storing a matching vector distance between the vector of the image of character to be identified and a vector of the matching character by a processor; and storing a data of the matching character according to the image of character to be identified when the matching vector distance is less than a vector distance threshold by the processor.

Character recognition method, character recognition device and non-transitory computer readable medium

A method for training a deep learning network for face recognition includes: utilizing a face landmark detector to perform face alignment processing on at least one captured image, thereby outputting at least one aligned image; inputting the at least one aligned image to a teacher model to obtain a first output vector; inputting the at least one captured image a student model corresponding to the teacher module to obtain a second output vector; and adjusting parameter settings of the student model according to the first output vector and the second output vector.

Face recognition network model with face alignment based on knowledge distillation

A method for image compression and a circuit system thereof are provided. In the method, pixel values of an image are obtained. A compression scenario is decided, for example, a uniform-quantization manner or a non-uniform-quantization manner is used for an M-bit image being compressed to an N-bit image so as to decide codeword sections for the image. Every codeword section has a codeword distance. The codeword sections have a fixed codeword distance in the uniform-quantization manner. Alternatively, in the non-uniform-quantization manner, the image can be divided into multiple codeword sections having different codeword distances according to a brightness distribution. Afterwards, a random number is generated for deciding codeword and index for original value of each of the pixels. An index table is accordingly formed. The index table is provided for obtaining the codeword in a decoding process by querying a codebook with the index.

Method for image compression and circuit system thereof

The present invention provides a processing circuit applied to a face recognition system, which includes a characteristic value calculation module, a determination circuit and a threshold value calculation module. The characteristic value calculation module is used to receive an image and process the image to generate a specific characteristic value; when the face recognition system operates in a face recognition phase, the determination circuit calculates multiple differences each between the specific characteristic value and one of multiple reference characteristic values, and determines whether at least one of the multiple differences is lower than a threshold value to generate a determination result; and when the face recognition system operates in a face registration phase, the threshold value calculation module determines a new threshold value according to differences between the specific characteristic value and the multiple reference values, for updating the threshold value used by the determination circuit.

Growing community of inventors

Hsinchu County, Taiwan

Shih-Tse Chen