Abstract: A method includes that a decoder side processes, based on a group of feature domain optical flows corresponding to an image frame, a first feature map of a reference frame to obtain a group of intermediate feature maps. The decoder side fuses the group of intermediate feature maps to obtain a predicted feature map, and the decoder side decodes the image frame based on the predicted feature map to obtain a target image. The predicted feature map of the image frame is determined by the decoder side by fusing a plurality of intermediate feature maps, and the predicted feature map includes more image information.

Abstract: A method includes determining a current feature and a reference feature. The current feature is of a to-be-encoded current image, and the reference feature is of a reference image of the to-be-encoded current image. A correlation matrix of the reference feature relative to the current feature is determined, and an inter motion feature is determined based on the correlation matrix. The inter motion feature is encoded into a bitstream.

Abstract: An encoding method includes obtaining a to-be-encoded frame, where the to-be-encoded frame is a P-frame, determining, from M preset network parameter sets, a network parameter set corresponding to the to-be-encoded frame, where the M preset network parameter sets respectively correspond to different compression performance information, and M is an integer greater than one, and encoding, by an encoding network, and based on the network parameter set corresponding to the to-be-encoded frame, the to-be-encoded frame to obtain a bitstream representative of the to-be-encoded frame.

Abstract: This application provides a feature domain optical flow determining method and a related device, and relates to the field of video or picture compression technologies based on artificial intelligence (AI). The method specifically includes: obtaining a picture domain optical flow between a current frame and a reference frame; performing multi-scale feature extraction on the reference frame, to obtain M feature maps of the reference frame, where M is an integer greater than or equal to 1; and performing M times of feature domain optical flow estimation based on the M feature maps of the reference frame and the picture domain optical flow between the current frame and the reference frame, to obtain M feature domain optical flows. A feature domain optical flow obtained by using the solutions of this application is more accurate and more stable, thereby improving inter-prediction accuracy.

Abstract: Embodiments of this application disclose a method and an apparatus for determining an image loss value, a storage medium, and a program product, and belong to the field of image compression technologies. In this method, loss values of different areas in an image are determined based on a partition indication map of the image, and then a total loss value is determined based on the loss values of the different areas. The partition indication map may be used to distinguish between a heavily-structured area and a lightly-structured area in the image, that is, the partition indication map may be used to distinguish between an edge structure and a texture. When the total loss value is used to assess image reconstruction quality, the image reconstruction quality can be assessed more comprehensively, and assessment of reconstruction quality of the edge structure and the texture can be maximally prevented from mutual impact.

Abstract: This application provides a feature map encoding and decoding method and an apparatus, and relates to the field of artificial intelligence (AI)-based data encoding and decoding technologies. The feature map decoding method includes: obtaining a bitstream of a to-be-decoded feature map, where the to-be-decoded feature map includes a plurality of feature elements; obtaining a first probability estimation result corresponding to each feature element based on the bitstream, where the first probability estimation result includes a first peak probability; determining a set of first feature elements and a set of second feature elements from the plurality of feature elements based on a first threshold and the first peak probability corresponding to each feature element; and obtaining a decoded feature map based on the set of first feature elements and the set of second feature elements. This can improve encoding and decoding performance while reducing encoding and decoding complexity.

Abstract: This application provides picture or audio encoding and decoding methods and apparatuses, and relates to the field of artificial intelligence (AI)—based picture or audio encoding and decoding technologies, and specifically, to the field of neural network-based picture feature map or audio feature variable encoding and decoding technologies. The encoding method includes: obtaining a to-be-encoded target, where the to-be-encoded target includes a plurality of feature elements, and the plurality of feature elements include a first feature element. The method further includes: obtaining a probability estimation result of the first feature element; determining, based on the probability estimation result of the first feature element, whether to perform entropy encoding on the first feature element; and performing entropy encoding on the first feature element only when it is determined that entropy encoding needs to be performed on the first feature element.

Abstract: Embodiments of this application disclose encoding and decoding methods, apparatuses, and devices, a storage medium, and a computer program, which relate to the field of encoding and decoding technologies. In embodiments of this application, in a decoding process, a plurality of feature points are divided into a plurality of groups based on a specified numerical value, and probability distributions of feature points in a same group are determined in parallel to improve decoding efficiency. Correspondingly, in an encoding process, the plurality of feature points are also grouped in a same grouping manner, and first image features of each group of feature points in the plurality of groups are sequentially encoded into a bit stream. To be concise, this solution can break through an efficiency bottleneck caused by serial computing when decoding is performed based on a variational auto-encoder (VAE), thereby effectively improving decoding efficiency.

Abstract: This application provides an audio/video or picture compression method and apparatus, which relates to the field of artificial intelligence (AI)-based audio/video or picture compression technologies, and to the field of neural network-based audio/video or picture compression technologies. The method includes: transforming a raw audio/video or picture to feature space through a multilayer convolution operation, extracting features of different layers in the feature space, outputting rounded feature signals of the different layers, predicting probability distribution of shallow feature signals by using deep feature signals or entropy estimation results, and performing entropy encoding on the rounded feature signals. In this application, signal correlation between different layers is utilized. In this way, audio/video or picture compression performance can be improved.

Abstract: A video frame compression method includes determining a target neural network from a plurality of neural networks according to a network selection policy; and generating, by using the target neural network, compression information corresponding to a current video frame. If the compression information is obtained by using a first neural network, the compression information includes first compression information of a first feature of the current video frame, and a reference frame of the current video frame is used for a compression process of the first feature of the current video frame. If the compression information is obtained by using a second neural network, the compression information includes second compression information of a second feature of the current video frame, and a reference frame of the current video frame is used for a generation process of the second feature of the current video frame.