-
公开(公告)号:US20220014767A1
公开(公告)日:2022-01-13
申请号:US17296432
申请日:2019-12-03
Applicant: ARM LIMITED
Inventor: Dominic Hugo SYMES , Sven Ola Johannes HUGOSSON
IPC: H04N19/34 , H04N19/176
Abstract: Disclosed herein is a method of encoding an array of data elements comprising transforming the array from the spatial to the frequency domain, representing the frequency domain coefficients as a plurality of bit plane arrays, and encoding the set of frequency domain coefficients as a data packet having a fixed size by encoding the bit plane arrays in a bit plane sequence working from the bit plane array representing the most significant bit downwards until the data packet is full. Each bit plane array is encoded by recursively subdividing the bit plane array into respective sections and subsections down to the individual coefficients and including in the data packet, so long as there is available space, data indicating the locations of any (sub)sections in that bit plane array that for the first time in the bit plane sequence contain one or more coefficient(s) having a non-zero bit value.
-
公开(公告)号:US20210350240A1
公开(公告)日:2021-11-11
申请号:US16871785
申请日:2020-05-11
Applicant: Arm Limited
Inventor: Erik PERSSON , Sven Ola Johannes HUGOSSON
Abstract: A method for adapting a trained neural network is provided. Input data is input to the trained neural network and a plurality of filters are applied to generate a plurality of channels of activation data. Differences between corresponding activation values in the plurality of channels of activation data are calculated and an order of the plurality of channels is determined based on the calculated differences. The neural network is adapted so that it will output channels of activation data in the determined order. The ordering of the channels of activation data is subsequently used to compress activation data values by taking advantage of a correlation between activation data values in adjacent channels.
-