A multi-task identification method, a training method, an electronic device, and a storage medium are provided, which relate to a field of an artificial intelligence technology, in particular to fields of deep learning, image processing and computer vision technologies, and may be applied to scenarios such as human faces. A specific implementation solution includes: obtaining first intermediate feature data according to an image to be identified; selecting a feature extraction strategy having a greatest matching degree with the image to be identified from a plurality of feature extraction strategies based on a target selection strategy and the first intermediate feature data, so as to obtain a target feature extraction strategy; processing the first intermediate feature data based on the target feature extraction strategy, to obtain second intermediate feature data; and obtaining a multi-task identification result for the image to be identified according to the second intermediate feature data.

A multi-task identification method, a training method, an electronic device, and a storage medium are provided, which relate to a field of an artificial intelligence technology, in particular to fields of deep learning, image processing and computer vision technologies, and may be applied to scenarios such as human faces. A specific implementation solution includes: obtaining first intermediate feature data according to an image to be identified; selecting a feature extraction strategy having a greatest matching degree with the image to be identified from a plurality of feature extraction strategies based on a target selection strategy and the first intermediate feature data, so as to obtain a target feature extraction strategy; processing the first intermediate feature data based on the target feature extraction strategy, to obtain second intermediate feature data; and obtaining a multi-task identification result for the image to be identified according to the second intermediate feature data.

Techniques of performing a visual search include updating probability distributions based on a succession of frames containing object images until a specified condition has been satisfied and producing a search result for the object only after the specified condition has been satisfied. When a user captures an image of a scene using a device, a front-end, visual search application running on the device obtains successive image frames and sends a first image frame to a back-end computer configured to perform a classification on the frame. The back-end computer obtains a prior probability distribution and generates a likelihood function indicating whether the image frame includes an object. The back-end computer then updates the prior probability distribution by adding respective values of parameters associated with the prior and likelihood function.

Techniques of performing a visual search include updating probability distributions based on a succession of frames containing object images until a specified condition has been satisfied and producing a search result for the object only after the specified condition has been satisfied. When a user captures an image of a scene using a device, a front-end, visual search application running on the device obtains successive image frames and sends a first image frame to a back-end computer configured to perform a classification on the frame. The back-end computer obtains a prior probability distribution and generates a likelihood function indicating whether the image frame includes an object. The back-end computer then updates the prior probability distribution by adding respective values of parameters associated with the prior and likelihood function.

Techniques for increasing robustness of a convolutional neural network based on training that uses multiple datasets and multiple tasks are described. For example, a computer system trains the convolutional neural network across multiple datasets and multiple tasks. The convolutional neural network is configured for learning features from images and accordingly generating feature vectors. By using multiple datasets and multiple tasks, the robustness of the convolutional neural network is increased. A feature vector of an image is used to apply an image-related operation to the image. For example, the image is classified, indexed, or objects in the image are tagged based on the feature vector. Because the robustness is increased, the accuracy of the generating feature vectors is also increased. Hence, the overall quality of an image service is enhanced, where the image service relies on the image-related operation.

The present invention in some embodiments thereof relates to a system and method for detecting inappropriate content on a device and filtering content on a variety of media. Inappropriate content is detected by taking a sample of media from at least one of a local memory, a data stream from a network and a data stream from local sensor, preprocessing the sample using a local processor and locally stored software to determine if the sample is a likely candidate to include objectionable content, in response to said sample being found to be a likely candidate perform at least one of quarantining the sample, marking the media, sending the sample to a remote processor for further analysis, analyzing the sample using an artificial intelligence routine running on a local processor and analyzing the sample using an artificial intelligence routine running on said local processor.

The present invention in some embodiments thereof relates to a system and method for detecting inappropriate content on a device and filtering content on a variety of media. Inappropriate content is detected by taking a sample of media from at least one of a local memory, a data stream from a network and a data stream from local sensor, preprocessing the sample using a local processor and locally stored software to determine if the sample is a likely candidate to include objectionable content, in response to said sample being found to be a likely candidate perform at least one of quarantining the sample, marking the media, sending the sample to a remote processor for further analysis, analyzing the sample using an artificial intelligence routine running on a local processor and analyzing the sample using an artificial intelligence routine running on said local processor.

A method for determining a salient object in an image based on superpixel analysis consists of initializing superpixels using SLIC algorithm and merging of adjacent superpixels that have similar color distribution. Then, calculate the spatial and color distribution correlation between the superpixels after being merged. Combined with the statistics of the occupancy rate, the distance to the original image center, and the global contrast of the superpixels, calculate the saliency evaluation vector for each superpixel. Finally, interpolate the saliency for each pixel in the superpixel.

A method for determining a salient object in an image based on superpixel analysis consists of initializing superpixels using SLIC algorithm and merging of adjacent superpixels that have similar color distribution. Then, calculate the spatial and color distribution correlation between the superpixels after being merged. Combined with the statistics of the occupancy rate, the distance to the original image center, and the global contrast of the superpixels, calculate the saliency evaluation vector for each superpixel. Finally, interpolate the saliency for each pixel in the superpixel.

A method of classifying motion from video frames involves generating motion frames indicative of changes in pixel values between pairs of frames. The method also involves determining one-dimensional feature values based on the video frames or motion frames, such as the statistical values or linear transformation coefficients. Each one-dimensional feature value may be stored in a buffer, from which additional temporal feature values can be extracted indicative of the change of the one-dimensional feature values across a set of frames. A classifier may receive the one-dimensional feature values and the additional temporal feature values as inputs, and determine the class of motion present in the video frames. Some classes of motion, such as irrelevant motion, may be considered irrelevant to the execution of certain motion-triggered actions, such that the method may involve suppressing the performance of a motion-triggered action based on the determined class of motion.