# US 2021/0200993 A1 · Condense-expansion-depth-wise convolutional neural network for face recognition > Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt - Number: US 2021/0200993 A1 - Document ID: US20210200993A1 - Status: Published application - Kind: A1 - Application number: 17/058,078 - Filed: 2018-09-13 - Published: 2021-07-01 - Earliest claimed: 2018-09-13 - Assignees: Intel Corporation - Applicants: Intel Corporation - Inventors: Yurong CHEN; Jianguo LI - Web page: https://patentgrep.com/patent/US20210200993A1 ## CPC - G06N 3/063: using electronic means - G06N 3/04: Architecture, e.g. interconnection topology - G06N 3/045: Combinations of networks - G06N 3/0464: Convolutional networks [CNN, ConvNet] - G06N 3/0495: Quantised networks; Sparse networks; Compressed networks - G06N 3/08: Learning methods - G06N 3/084: Backpropagation, e.g. using gradient descent - G06N 3/09: Supervised learning - G06V 10/32: Normalisation of the pattern dimensions - G06V 10/454: {Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]} - G06V 10/82: using neural networks - G06V 40/161: {Detection; Localisation; Normalisation} - G06V 40/168: {Feature extraction; Face representation} - G06V 40/172: {Classification, e.g. identification} - G06N 3/048: Activation functions - G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS - G06V: IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING ## Abstract Techniques related to implementing convolutional neural networks for face or other object recognition are discussed. Such techniques may include applying, in turn, a depth-wise separable convolution, a condense point-wise convolution, and an expansion point-wise convolution to input feature maps to generate output feature maps such that the output from the expansion point-wise convolution has more channels than the output from the condense point-wise convolution. ## Claims ### Claim at position 1 1-25. (canceled) ### Claim 26 A system for implementing a convolutional neural network (CNN) comprising: a memory to store input image data; and a processor coupled to the memory, the processor to: apply a depth-wise separable convolution to a plurality of input feature maps to generate a plurality of separate 2D feature maps, wherein the plurality of input feature maps are associated with the input image data; apply a condense point-wise convolution to the plurality of separate 2D feature maps to generate a first plurality of combined feature maps, wherein the first plurality of combined feature maps has a first number of channels; apply an expansion point-wise convolution to the first plurality of combined feature maps to generate a second plurality of combined feature maps, wherein the second plurality of combined feature maps has a second number of channels greater than the first number of channels; and output object recognition data corresponding to the input image data based at least in part on the second plurality of combined feature maps. ### Claim 27 (depends on 26) The system of claim 26, wherein the plurality of input feature maps has a number of channels that matches the second number. ### Claim 28 (depends on 27) The system of claim 27, the processor to: perform a residual connection by summation of the plurality of input feature maps and the second plurality of combined feature maps to generate a plurality of output feature maps; and provide the plurality of output feature maps or a second plurality of output feature maps corresponding to the plurality of output feature maps to a second depth-wise separable convolution of the CNN. ### Claim 29 (depends on 28) The system of claim 28, the processor to: apply, in turn, the second depth-wise separable convolution, a second condense point-wise convolution, and a second expansion point-wise convolution to the plurality of output feature maps or the second plurality of output feature maps to generate a third plurality of output feature maps, wherein the third plurality of output feature maps has a third number of channels greater than the second number of channels. ### Claim 30 (depends on 29) The system of claim 29, wherein the second condense point-wise convolution generates a third plurality of combined feature maps having a third number of channels, the second expansion point-wise convolution generates a fourth plurality of combined feature maps having a fourth number of channels, and a ratio of the second number of channels to the first number of channels is the same as a ratio of the fourth number of channels to the third number of channels. ### Claim 31 (depends on 26) The system of claim 26, the processor to: apply, in turn, a second depth-wise separable convolution, a second condense point-wise convolution, and a second expansion point-wise convolution to the second plurality of combined feature maps to generate a third plurality of combined feature maps, wherein the third plurality of output feature maps has a third number of channels equal to the second number of channels. ### Claim 32 (depends on 26) The system of claim 26, wherein the processor to apply the depth-wise separable convolution comprises the processor to apply the second number, n, of k×k×1 kernels to each of the plurality of input feature maps, the processor to apply the condense point-wise convolution comprises the processor to apply the first number, n/g, of 1×1×n kernels to the plurality of separate 2D feature maps, and the processor to apply the expansion point-wise convolution comprises the processor to apply the second number, n, of 1×1×(n/g) kernels to the plurality of separate 2D feature maps. ### Claim 33 (depends on 32) The system of claim 32, wherein a ratio of the second number, n, of channels to the first number, n/g, of channels comprises a condense factor, g, that is greater than two. ### Claim 34 (depends on 26) The system of claim 26, wherein the input image data comprises one of RGB image data, RGB-D image data, or infra-red image data. ### Claim 35 (depends on 26) The system of claim 26, wherein the object recognition data comprises one of an indicator of whether the input image data corresponds to a face of a user or a label corresponding to one of a plurality of candidate faces. ### Claim 36 A computer-implemented method for implementing a convolutional neural network (CNN) comprising: applying a depth-wise separable convolution to a plurality of input feature maps to generate a plurality of separate 2D feature maps, wherein the plurality of input feature maps are associated with input image data; applying a condense point-wise convolution to the plurality of separate 2D feature maps to generate a first plurality of combined feature maps, wherein the first plurality of combined feature maps has a first number of channels; applying an expansion point-wise convolution to the first plurality of combined feature maps to generate a second plurality of combined feature maps, wherein the second plurality of combined feature maps has a second number of channels greater than the first number of channels; and outputting object recognition data corresponding to the input image data based at least in part on the second plurality of combined feature maps. ### Claim 37 (depends on 36) The method of claim 36, wherein the plurality of input feature maps has a number of channels that matches the second number. ### Claim 38 (depends on 37) The method of claim 37, further comprising: performing a residual connection by summation of the plurality of input feature maps and the second plurality of combined feature maps to generate a plurality of output feature maps; and providing the plurality of output feature maps or a second plurality of output feature maps corresponding to the plurality of output feature maps to a second depth-wise separable convolution of the CNN. ### Claim 39 (depends on 38) The method of claim 38, further comprising: applying, in turn, the second depth-wise separable convolution, a second condense point-wise convolution, and a second expansion point-wise convolution to the plurality of output feature maps or the second plurality of output feature maps to generate a third plurality of output feature maps, wherein the third plurality of output feature maps has a third number of channels greater than the second number of channels. ### Claim 40 (depends on 39) The method of claim 39, wherein the second condense point-wise convolution generates a third plurality of combined feature maps having a third number of channels, the second expansion point-wise convolution generates a fourth plurality of combined feature maps having a fourth number of channels, and a ratio of the second number of channels to the first number of channels is the same as a ratio of the fourth number of channels to the third number of channels. ### Claim 41 At least one machine readable medium comprising a plurality of instructions that, in response to being executed on a computing device, cause the computing device to implement a convolutional neural network (CNN) by: applying a depth-wise separable convolution to a plurality of input feature maps to generate a plurality of separate 2D feature maps, wherein the plurality of input feature maps are associated with input image data; applying a condense point-wise convolution to the plurality of separate 2D feature maps to generate a first plurality of combined feature maps, wherein the first plurality of combined feature maps has a first number of channels; applying an expansion point-wise convolution to the first plurality of combined feature maps to generate a second plurality of combined feature maps, wherein the second plurality of combined feature maps has a second number of channels greater than the first number of channels; and outputting object recognition data corresponding to the input image data based at least in part on the second plurality of combined feature maps. ### Claim 42 (depends on 41) The machine readable medium of claim 41, wherein the plurality of input feature maps has a number of channels that matches the second number. ### Claim 43 (depends on 42) The machine readable medium of claim 42, further comprising instructions that, in response to being executed on the computing device, cause the computing device to implement the convolutional neural network (CNN) by: performing a residual connection by summation of the plurality of input feature maps and the second plurality of combined feature maps to generate a plurality of output feature maps; and providing the plurality of output feature maps or a second plurality of output feature maps corresponding to the plurality of output feature maps to a second depth-wise separable convolution of the CNN. ### Claim 44 (depends on 43) The machine readable medium of claim 43, further comprising instructions that, in response to being executed on the computing device, cause the computing device to implement the convolutional neural network (CNN) by: applying, in turn, the second depth-wise separable convolution, a second condense point-wise convolution, and a second expansion point-wise convolution to the plurality of output feature maps or the second plurality of output feature maps to generate a third plurality of output feature maps, wherein the third plurality of output feature maps has a third number of channels greater than the second number of channels. ### Claim 45 (depends on 44) The machine readable medium of claim 44, wherein the second condense point-wise convolution generates a third plurality of combined feature maps having a third number of channels, the second expansion point-wise convolution generates a fourth plurality of combined feature maps having a fourth number of channels, and a ratio of the second number of channels to the first number of channels is the same as a ratio of the fourth number of channels to the third number of channels.