US 12,094,084 B2
Multi-channel feature map fusion
Bin Sun, Shenzhen (CN); Mingguo Zhao, Shenzhen (CN); and Youjun Xiong, Shenzhen (CN)
Assigned to UBTECH ROBOTICS CORP LTD, Shenzhen (CN)
Filed by UBTECH ROBOTICS CORP LTD, Shenzhen (CN)
Filed on Jul. 29, 2021, as Appl. No. 17/388,043.
Application 17/388,043 is a continuation of application No. PCT/CN2020/140923, filed on Dec. 29, 2020.
Claims priority of application No. 202010775325.2 (CN), filed on Aug. 5, 2020.
Prior Publication US 2022/0044370 A1, Feb. 10, 2022
Int. Cl. G06T 5/20 (2006.01); G06F 7/24 (2006.01); G06T 5/50 (2006.01)
CPC G06T 5/20 (2013.01) [G06F 7/24 (2013.01); G06T 5/50 (2013.01); G06T 2207/20021 (2013.01); G06T 2207/20221 (2013.01)] 3 Claims
OG exemplary drawing
 
1. An image processing method, comprising steps of:
obtaining to-be-processed multi-channel feature maps, wherein each of the multi-channel feature maps corresponding to a channel;
obtaining multi-channel first output feature maps by processing the multi-channel feature maps through a pointwise convolution and multi-channel second output feature maps by processing the multi-channel feature maps through a non-pointwise convolution, wherein the pointwise convolution and the non-pointwise convolution are performed in parallel, and the non-pointwise convolution is for describing a spatial feature of each channel and an information exchange between the channel feature maps; and
fusing the multi-channel first output feature maps and the multi-channel second output feature maps to obtain multi-channel third output feature maps;
wherein, the non-pointwise convolution is an adaptive linear operation, and in response to a first ratio of a channel amount corresponding to the channel feature maps to perform the non-pointwise convolution and a channel amount corresponding to the multi-channel second output feature maps being a positive integer, the multi-channel second output feature maps is obtained by:
evenly dividing the channel feature maps into a plurality of feature map groups by channel according to the first ratio, wherein an amount of the feature maps in each feature map group is the first ratio; and
performing the linear operation on the feature map corresponding to each channel in each feature map group, and using a cumulative sum of results of the linear operations of the feature map group as the second output feature maps corresponding to the feature map group.