Abstract: Methods and systems for updating the weights of a set of convolution kernels of a convolutional layer of a neural network are described. A set of convolution kernels having attention-infused weights is generated by using an attention mechanism based on characteristics of the weights. For example, a set of location-based attention multipliers is applied to weights in the set of convolution kernels, a magnitude-based attention function is applied to the weights in the set of convolution kernels, or both. An output activation map is generated using the set of convolution kernels with attention-infused weights. A loss for the neural network is computed, and the gradient is back propagated to update the attention-infused weights of the convolution kernels.

Abstract: In some examples, a system balances a number of positive data points and a number of negative data points, to produce a balanced training data set, where the positive data points comprise features associated with authentication events that are positive with respect to an unauthorized classification, and the negative data points comprise features associated with authentication events that are negative with respect to the unauthorized classification. The system trains a plurality of models using the balanced training data set, wherein the plurality of models are trained according to respective different machine learning techniques. The system selects a model from the trained plurality of models based on relative performance of the plurality of models.

Abstract: Disclosed are a liveness determining method and apparatus and a method and apparatus for training the liveness determining apparatus. The liveness determining method includes extracting, by a processor, a feature from an input fingerprint image, inputting the feature into the current layer classifier, inputting the feature into the subsequent layer classifier, based on a determination that an output of the current layer classifier is live, and determining a liveness of the input fingerprint image to be false, based on a determination that an output of the subsequent layer classifier is fake, wherein the current layer classifier and the subsequent layer classifier are respectively trained based on a plurality of training fake images belonging to different groups.

Abstract: The present disclosure is directed to extracting text from form-like documents. In particular, a computing system can obtain an image of a document that contains a plurality of portions of text. The computing system can extract one or more candidate text portions for each field type included in a target schema. The computing system can generate a respective input feature vector for each candidate for the field type. The computing system can generate a respective candidate embedding for the candidate text portion. The computing system can determine a respective score for each candidate text portion for the field type based at least in part on the respective candidate embedding for the candidate text portion. The computing system can assign one or more of the candidate text portions to the field type based on the respective scores.

Abstract: Training a network for image processing with temporal consistency includes obtaining un-annotated frames from a video feed. A pretrained network is applied to the first frame of first frame set comprising a plurality of frames to obtain a first prediction, wherein the pretrained network is pretrained for a first image processing task. A current version of the pretrained network is applied to each frame of the first frame set to obtain a first prediction. A content loss term is determined, based on the first prediction and a current prediction for the frame, based on the current network. A temporal consistency loss term is also determined based on a determined consistency of pixels within each frame of the first frame set. The pretrained network may be refined based on the content loss term and the temporal term to obtain a refined network.

Abstract: A recognition apparatus for extracting first features of an object from a region of the object; obtaining second features of the object at least based on confidence information for first attribute of the object; determining sample information from pre-determined sample information based on the obtained second features; and recognizing the first attribute of the object based on the determined sample information.

Abstract: A monitoring camera having artificial intelligence includes an imaging unit, a communication unit that receives a parameter relating to a detection target from a terminal device, and a processing unit that constructs the artificial intelligence based on the parameter, and uses the constructed artificial intelligence to detect the detection target from an image captured by the imaging unit.

Abstract: The method includes for each of a plurality of successive images of a video stream from a camera, the search for at least one person present in the image and the definition, for each person found, of a zone in the image, known as person zone, surrounding this person at least partially; for each of at least one person, the grouping together into one tracklet of several person zones from successive images and surrounding this same person at least partially; for each tracklet: the identification of the person in this tracklet from person zones, the determination of a moment at which the line is crossed by the person identified from person zones and the addition of the name found and of the moment of crossing determined in at least some of the images containing the person zones.

Abstract: Described are systems and methods for detecting an anomaly among a plurality of components operating in a system. In some embodiments, a method includes monitoring a plurality of metrics for the plurality of components across a plurality of time periods. For each time period, the plurality of components is clustered into a plurality of clusters based on measurement information corresponding to the plurality of metrics received for the time period. For each component, a plurality of correspondences is determined between the clusters in which the component is grouped for a plurality of pairs of adjacent time periods. Then, whether each component is operating anomalously can be determined based on the plurality of determined correspondences.

Abstract: An evaluation assistance method includes: acquiring a first image to be used for performance evaluation of trained models; generating a plurality of second images, each of the plurality of second images being a result of processing the first image by each of a plurality of trained models; and displaying each of the plurality of trained models in association with a corresponding second image of the plurality of second images.

Abstract: Techniques for a framework for building, orchestrating, and deploying complex, large-scale Machine Learning (ML) or deep learning (DL) inference applications is described. A ML application orchestration service is disclosed that enables the construction, orchestration, and deployment of complex ML inference applications in a provider network. The disclosed service provides customers with the ability to define machine learning (ML) models and define transformation operations on data before and/or after being provided to the ML models to construct a complex ML inference application. The service provides a framework for the orchestration (co-ordination) of the workflow logic (e.g., of the request and/or response flows) involved in building and deploying a complex ML inference application in the provider network.

Abstract: Object re-identification refers to a process by which images that contain an object of interest are retrieved from a set of images captured using disparate cameras or in disparate environments. Object re-identification has many useful applications, particularly as it is applied to people (e.g. person tracking). Current re-identification processes rely on convolutional neural networks (CNNs) that learn re-identification for a particular object class from labeled training data specific to a certain domain (e.g. environment), but that do not apply well in other domains. The present disclosure provides cross-domain disentanglement of id-related and id-unrelated factors. In particular, the disentanglement is performed using a labeled image set and an unlabeled image set, respectively captured from different domains but for a same object class.

Abstract: A computer-implemented method for emulating an object recognizer includes receiving testing image data, and emulating, by employing a first object recognizer, a second object recognizer. Emulating the second object recognizer includes using the first object recognizer to perform object recognition on a testing object from the testing image data to generate data, the data including a feature representation for the testing object, and classifying the testing object based on the feature representation and a machine learning model configured to predict whether the testing object would be recognized by a second object recognizer. The method further includes triggering an action to be performed based on the classification.

Abstract: There is provided a system and method for optimizing depth imaging. The method including: illuminating one or more scenes with illumination patterns; capturing one or more images of each of the scenes; reconstructing the scenes; estimating the reconstruction error and a gradient of the reconstruction error; iteratively performing until the reconstruction error reaches a predetermined error condition: determining a current set of control vectors and current set of reconstruction parameters; illuminating the one or more scenes with the illumination patterns governed by the current set of control vectors; capturing one or more images of each of the scenes while the scene is being illuminated with at least one of the illumination patterns; reconstructing the scenes from the one or more captured images using the current reconstruction parameters; and estimating an updated reconstruction error and gradient; and outputting at least one of control vectors and reconstruction parameters.

Abstract: A cross-modality processing method is related to a field of natural language processing technologies. The method includes: obtaining a sample set, wherein the sample set includes a plurality of corpus and a plurality of images; generating a plurality of training samples according to the sample set, in which each of the plurality of the training samples is a combination of at least one of the plurality of the corpus and at least one of the plurality of the images corresponding to the at least one of the plurality of the corpus; adopting the plurality of the training samples to train a semantic model, so that the semantic model learns semantic vectors containing combinations of the corpus and the images.

Abstract: Systems and methods for determining a continuous grain size log from a collection of petrographic thin section images are provided. Thin section images from core samples from one or more wells may be obtained and analyzed to estimate grain sizes. Using wireline logs from the one or more wells and the estimated grain sizes, a data structure (for example, a database) of grain sizes and wireline logs at depths may be constructed. The data structure may be used to train a machine learning model. Next, a wireline tool may be used to obtain wireline logs in a new well, and a continuous grain size log may be determined from the wireline logs of using the machine learning model. Computer-readable media for determining reservoir rock grain sizes from a collection of petrographic thin section images is also provided.

Abstract: A system includes a computing device that includes a memory configured to store instructions. The system also includes a processor to execute the instructions to perform operations that include determining a ranking of images using a machine learning system. The machine learning system is trained using attributes that represent each of a plurality of training images. The attributes include imagery attributes, social network attributes, and textual attributes. Operations also include producing a listing of the ranked images for selecting one or more of the ranked images for a brand entity associated with the selected ranked images.

Abstract: A label feature extraction means 71 extracts, from reference information, a label feature that is a vector representing a feature of the reference information. A label feature dimension reduction means 72 performs dimension reduction of the label feature. An image feature extraction means 73 extracts an image feature from a target image that is an image in which an object to be recognized is captured. A feature transformation means 74 performs feature transformation on the image feature in such a manner that comparison with the label feature after the dimension reduction becomes possible. The class recognition means 75 recognizes a class of the object to be recognized by comparing the image feature after the feature transformation with the label feature after the dimension reduction.

Abstract: A parking assist apparatus includes: a map generation unit that generates an obstacle map showing a position of an obstacle existing in a surrounding of own vehicle based on a detected result of a sonar, and records data indicating existence of the obstacle on the obstacle map; a wheel-stop detection unit that divides the obstacle map generated by the map generation unit into a plurality of areas, counts a number of data pieces in each of the plurality of divided areas, and detects a position of a wheel stop based on the counted number of data pieces in each of the plurality of areas; and a parking position setting unit that sets a parking position for parking the own vehicle based on the position of the wheel stop detected by the wheel-stop detection unit.

Abstract: The technology disclosed relates to authenticating users using a plurality of non-deterministic registration biometric inputs. During registration, a plurality of non-deterministic biometric inputs are given as input to a trained machine learning model to generate sets of feature vectors. The non-deterministic biometric inputs can include a plurality of face images and a plurality of voice samples of a user. A characteristic identity vector for the user can be determined by averaging feature vectors. During authentication, a plurality of non-deterministic biometric inputs are given as input to a trained machine learning model to generate a set of authentication feature vectors. The sets of feature vectors are projected onto a surface of a hyper-sphere. The system can authenticate the user when a cosine distance between the authentication feature vector and a characteristic identity vector for the user is less than a pre-determined threshold.

Abstract: The present disclosure provides a visual localization method and apparatus based on a semantic error image.

Abstract: There are provided a classifier and method of classifying defects in a semiconductor specimen. The method comprises receiving defects classified into a majority class, each having values for plurality of attributes, some defects belonging to a minority class, and some to the majority; selecting an attribute subset and defining differentiators for attributes wherein a second classifier using the subset and differentiators classifies correctly to minority and majority classes at least part of the defects; generating a training set comprising: defects of the majority and minority classes, and additional defects which the second classifier classifies as minority; training, upon the training set, subset, and differentiators, an engine obtaining a confidence level that a defect belongs to the majority class; applying the engine to second defects classified to the majority class, to obtain a confidence level of classifying each defect to the majority class; and outputting defects having a low confidence level.

Abstract: System/method identifying a defined object (e.g., hazard): a sensor detecting and defining a digital representation of an object; a processor (connected to the sensor) which executes two techniques to identify a signature of the defined object; a memory (connected to the processor) storing reference data relating to two signatures derived, respectively, by the two techniques; responsive to the processor receiving the digital representation from the sensor, the processor executes the two techniques, each technique assessing the digital representation to identify any signature candidate defined by the object, derive feature data from each identified signature candidate, compare the feature data to the reference data, and derive a likelihood value of the signature candidate corresponding with the respective signature; combining likelihood values to derive a composite likelihood value and thus determine whether the object in the digital representation is the defined object.

Abstract: A hybrid data processing environment comprising a classical computing system and a quantum computing system is configured. A configuration of a first quantum circuit is produced from the classical computing system, the first quantum circuit being executable using the quantum computing system. Using the quantum computing system, the first quantum circuit is executed. Using a pattern recognition technique, a portion of the first quantum circuit that can be transformed using a first transformation operation to satisfy a constraint on the quantum circuit design is identified. The portion is transformed to a second quantum circuit according to the first transformation operation, wherein the first transformation operation comprises reconfiguring a gate in the first quantum circuit such that a qubit used in the gate complies with the constraint on the quantum circuit design while participating in the second quantum circuit. Using the quantum computing system, the second quantum circuit is executed.

Abstract: Devices, systems, and methods of evaluating risk associated with a condition of the vessel and providing an objective intervention recommendation based on the evaluated risk are disclosed. The method includes steps of obtaining physiologic measurements from a first instrument and a second instrument positioned within the vessel of the patient while the second instrument is moved longitudinally through the vessel from a first position to a second position, obtaining image data from an image of a vessel system, co-registering the physiologic measurements with the image data to produce co-registered physiologic measurements, and determining whether to perform a first surgical procedure or a second surgical procedure, wherein the determining is based on the co-registered physiologic measurements. Other associated methods, systems, and devices are also provided herein.

Abstract: One embodiment of the present invention sets forth a technique for evaluating labeled data. The technique includes selecting, from a set of labels for a data sample, a subset of the labels representing non-outliers in a distribution of values in the set of labels. The technique also includes aggregating the subset of the labels into a benchmark for the data sample. The technique further includes generating, based on a comparison between the benchmark and an additional label, a benchmark score associated with the data sample, and generating a set of performance metrics for labeling the data sample based on the benchmark score.

Abstract: An information management apparatus comprises a communication unit configured to communicate with a plurality of external apparatuses having learning functions, and a control unit configured to control the communication with the plurality of external apparatuses performed by the communication unit. The control unit, if supervisory data generated when a predetermined external apparatus executes a learning function is received from the predetermined external apparatus via the communication unit, selects, from among the plurality of external apparatuses, an external apparatus, other than the predetermined external apparatus, with which the supervisory data is to be shared, and performs control so that the supervisory data is transmitted to the selected external apparatus.

Abstract: A human expert may initially label a white light image of teeth, and computer vision may initially label a filtered fluorescent image of the same teeth. Each label may indicate presence or absence of dental plaque at a pixel. The images may be registered. For each pixel of the registered images, a union label may be calculated, which is the union of the expert label and computer vision label. The union labels may be applied to the white light image. This process may be repeated to create a training set of union-labeled white light images. A classifier may be trained on this training set. Once trained, the classifier may classify a previously unseen white light image, by predicting union labels for that image. Alternatively, the items that are initially labeled may comprise images captured by two different imaging modalities, or may comprise different types of sensor measurements.

Abstract: Computerized techniques for improved binarization and extraction of information from digital image data are disclosed in accordance with various embodiments. The inventive concepts include rendering a digital image using a plurality of binarization thresholds to generate a plurality of binarized digital images, wherein at least some of the binarized digital images are generated using one or more binarization thresholds that are determined based on a priori knowledge regarding an object depicted in the digital image; identifying one or more connected components within the plurality of binarized digital images; and identifying one or more text regions within the digital image based on some or all of the connected components. Systems and computer program products are also disclosed.

Abstract: Methods, systems, and processes to optimize role level identification for computing resource allocation to perform security operations in networked computing environments. A role level classifier to process a training dataset that corresponds to a clean title is generated from a subset of entities associated with the clean title. An initial effective title determined by the role level classifier based on processing the training dataset is assigned to an entity. A new effective title based on feature differences between the initial effective title and the clean title is re-assigned to the entity. Performance of the generating, the assigning, and the re-assigning is repeated using the new effective title instead of the clean title.

Abstract: An image inspection computing device is provided. The device includes a memory device and at least one processor. The at least one processor is configured to receive at least one sample image of a first component, wherein the at least one sample image of the first component does not include defects, store, in the memory, the at least one sample image, and receive an input image of a second component. The at least one processor is also configured to generate an encoded array based on the input image of the second component, perform a stochastic data sampling process on the encoded array, generate a decoded array, and generate a reconstructed image of the second component, derived from the stochastic data sampling process and the decoded array. The at least one processor is further configured to produce a residual image, and identify defects in the second component.

Abstract: A multi-camera vision system and method of monitoring. In one embodiment imaging systems provide object classifications with cameras positioned to receive image data from a field of view to classify an object among multiple classifications. A control unit receives classification or position information of objects and (ii) displays an image corresponding to a classified object relative to the position of the structure. An embodiment of a related method monitors positions of an imaged object about a boundary by continually capturing at least first and second series of image frames, each series comprising different fields of view of a scene about the boundary, with some of the image frames in the first series covering a wide angle field of view and some of the image frames in the second series covering no more than a narrow angle field of view.

Abstract: An artificial intelligence neural network apparatus, comprising: a labeled learning database having data of a feature vector composed of N elements; a first feature vector image converter configured to visualize the data in the learning database to form an imaged learning feature vector image database; a deep-learned artificial intelligence neural network configured to use a learning feature vector image in the learning feature vector image database to perform an image classification operation; an inputter configured to receive a test image, and generate test data based on the feature vector; and a second feature vector image converter configured to visualize the test data and convert the visualized test data into a test feature vector image. The deep-learned artificial intelligence neural network is configured to determine a class of the test feature vector image.

Abstract: Systems and methods include processors for receiving training data for a user activity; receiving bias criteria; determining a set of model parameters for a machine learning model including: (1) applying the machine learning model to the training data; (2) generating model prediction errors; (3) generating a data selection vector to identify non-outlier target variables based on the model prediction errors; (4) utilizing the data selection vector to generate a non-outlier data set; (5) determining updated model parameters based on the non-outlier data set; and (6) repeating steps (1)-(5) until a censoring performance termination criterion is satisfied; training classifier model parameters for an outlier classifier machine learning model; applying the outlier classifier machine learning model to activity-related data to determine non-outlier activity-related data; and applying the machine learning model to the non-outlier activity-related data to predict future activity-related attributes for the user activity

Abstract: An electronic device includes circuitry that determines probability map information for a first image, based on application of a neural network model on the first image. The neural network model is trained to detect one or more objects based on a plurality of images associated with the one or more objects. The probability map information indicates a probability value for each pixel in the first image. A region corresponding to the one or more objects is detected in the first image based on the probability map information. A first set of sub-images is determined from the detected region, based on application of a stochastic optimization function on the probability map information. The one or more objects are detected from a second set of sub-images of the first set of sub-images, based on application of the neural network model on the second set of sub-images.

Abstract: A method for image analysis according to an embodiment may include generating a prediction result for an original image using a pre-trained image classification model, learning a plurality of masks using the original image, the prediction result, and the image classification model, and generating a map visualizing a importance of each area of the original image for the prediction result based on at least one of the plurality of masks.

Abstract: Test case selection methods are disclosed. A feature of a candidate test case and respective features of a set of test cases are extracted. The set of test cases is clustered into a plurality of clusters based on the respective features of the set of test cases. At least one cluster related to the candidate test case is determined from the plurality of clusters based on the feature of the candidate test case. At least one test case similar to the candidate test case is selected from a plurality of test cases included in the at least one cluster.

Abstract: A tactile perception system is provided. The tactile perception system includes a storage unit storing tactile data and feature information corresponding to the tactile data, a sensing unit sensing surface characteristics of an object to generate a sensing signal, an extraction unit extracting sensing information from the sensing signal generated by the sensing unit, and a matching unit extracting a piece of feature information, which is matched with the sensing information, from the feature information stored in the storage unit and extracting a piece of tactile data, which corresponds to the piece of feature information, from the tactile data stored in the storage unit.

Abstract: Image Signal Processing (ISP) optimization framework for computer vision applications is disclosed. The tuning of the ISP is performed automatically and presented as a nonlinear multi-objective optimization problem, followed by solving the problem using an evolutionary stochastic solver. An improved ISP of the embodiments of the invention includes at least features of search space reduction for reducing a number of ISP configurations, remapping the generated population to the reduced search space via mirroring, and global optimization function processing, which allow tuning all the blocks of the ISP at the same time instead of the prior art tuning of each ISP block separately. Also shown that an ISP tuned for image quality performs inferior compared with an ISP trained for a specific downstream image recognition task.

Abstract: A machine learning system includes a first determination model that determines whether an input image is a second domain image, and a second determination model that determines whether an extracted image obtained by extracting an area where an object is presented from the input image is from the second domain image. Either a pseudo second domain image or the second domain image is selected and input into the first determination model, and either a first extracted image in the pseudo second domain image or a second extracted image in the second domain image is selected and input into an image extracting unit. Learning of the first determination model is performed based on a first determination result, learning of the second determination model is performed based on a second determination result, and learning of a pseudo image generative model is performed based on the first and second determination results.

Abstract: Example methods and systems for training deep learning engines for radiotherapy treatment planning are provided. One example method may comprise: obtaining a set of training data that includes unlabeled training data and labeled training data; and configuring a deep learning engine to include (a) a primary network and (b) a deep supervision network that branches off from the primary network. The method may further comprise: training the deep learning engine to perform the radiotherapy treatment planning task by processing training data instance to generate (a) primary output data and (b) deep supervision output data; and updating weight data associated with at least some of the multiple processing layers based on the primary output data and/or the deep supervision output data. The deep supervision network may be pruned prior to applying the primary network to perform the radiotherapy treatment planning task for a patient.

Abstract: The present disclosure relates to systems, non-transitory computer-readable media, and methods for training and utilizing scale-diverse segmentation neural networks to analyze digital images at different scales and identify different target objects portrayed in the digital images. For example, in one or more embodiments, the disclosed systems analyze a digital image and corresponding user indicators (e.g., foreground indicators, background indicators, edge indicators, boundary region indicators, and/or voice indicators) at different scales utilizing a scale-diverse segmentation neural network. In particular, the disclosed systems can utilize the scale-diverse segmentation neural network to generate a plurality of semantically meaningful object segmentation outputs. Furthermore, the disclosed systems can provide the plurality of object segmentation outputs for display and selection to improve the efficiency and accuracy of identifying target objects and modifying the digital image.

Abstract: A system for classification of scholastic works includes a computing device configured to receive a first scholastic work, identify an author and a category of the first scholastic work, determine at least a work theme by receiving theme training data, the theme training data including a plurality of entries, each entry including a training textual element and a correlated theme, training a theme classifier as a function of the training data, and determining the at least a work theme as a function of the plurality of textual elements and the theme classifier, calculate a reliability quantifier as a function of the at least a theme, the author, and the category, select the scholastic work as a function of the reliability quantifier, derive, from the scholastic work, at least a correlation between a diagnostic element and a practice, and store the at least a correlation in an expert database.

Abstract: A computing device divides a training image into a plurality of training blocks, and the training image includes a training object. The computing device calculates, for each of the training blocks, a correct confidence score of the training object covering the training block according to an image-marking data and a confidence-score-translating function, and the image-marking data includes a piece of location information of the training object in the training image. Then, the computing device trains a deep-learning model with the training image, the correct confidence scores and the image-marking data to generate the object-detecting model.

Abstract: A neural network batch normalization optimization method includes: setting a first network layer in a neural network as a starting layer; sequentially obtaining initial bias values of different network layers backwards starting from the starting layer; calculating equivalent bias values of the different network layers; determining whether there is a target network layer, wherein a ratio of the equivalent bias value corresponding to a previous layer of a target network layer to the equivalent bias value corresponding to the target network layer is no less than a pre-set threshold value; and if the target network layer is present, setting the bias values of the different network layers between the starting layer and the previous layer of the target network layer to zero, and taking the equivalent bias value of the target network layer as a bias value of the target network layer.

Abstract: A computer implemented method for optical character recognition (OCR) of a character string in a text image. The method efficiently combines two different OCR engines with the computation that needs to be done by the second OCR engine depending on the results found by the first OCR engine. This method provides, in particular, a high speed and accurate results when the first OCR engine is fast and the second OCR engine is accurate. The combination is possible because the second OCR engine identifies each segment to be processed by the second OCR engine without needing to process all segments.

Abstract: A method and data summarization system for dynamically generating summarised content for visual and contextual text data, is disclosed. The method includes classifying plurality of dataset related to one or more domains based on datatype associated with each dataset. The datatype comprises text, numeric and visual data. Upon classification, one or more usable tokens are determined from the text data using a predefined token learning model. Further, one or more graphical parameters are determined from the visual data by using a pre-trained graphical model. Thereafter, based on the one or more usable tokens and the one or more graphical parameters, a summarized content is generated for the plurality of dataset.

Abstract: Compositing aware digital image search techniques and systems are described that leverage machine learning. In one example, a compositing aware image search system employs a two-stream convolutional neural network (CNN) to jointly learn feature embeddings from foreground digital images that capture a foreground object and background digital images that capture a background scene. In order to train models of the convolutional neural networks, triplets of training digital images are used. Each triplet may include a positive foreground digital image and a positive background digital image taken from the same digital image. The triplet also contains a negative foreground or background digital image that is dissimilar to the positive foreground or background digital image that is also included as part of the triplet.

Abstract: A method for recognizing a face from monitoring video data is disclosed. Two neural networks are used to compare and score high-dimensional face features of a face, and a K-neighbor algorithm and a screening mechanism with a Euclidean distance as a threshold are combined for face comparison recognition to obtain an accurate face recognition result. In addition, the present disclosure also performs further screening based on the time of video data and the frequency of face appearance, and finally obtains a recognition result, thereby obtaining a more accurate face recognition result. The present disclosure can perform relatively accurate face recognition on video data or picture data captured by a real-time or historical monitoring camera.

Abstract: Disclosed is a fast side-face interference resistant face detection method, in which a user selects an ordinary image, uses a deep neural network to extract image features, and then determines an exact location of a face. A training method for face detection uses a pure data-driven manner, uses an ordinary face image and a face boundary box as inputs, uses mirror symmetry and Gaussian filtering to perform data augmentation, and uses migration learning and hard example mining to enhance training effects. After a face image is read, the image is firstly scaled, and then placed into the deep neural network to extract features, and generate a plurality of face likelihood boxes and confidence scores of the face likelihood boxes, and finally the most appropriate face likelihood box is selected in a non-maximum suppression manner. No specific requirements are set on an angle of the face image, and a detection effect of a side face is still very obvious.