Separable Convolutional Layers for Efficient Facial Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional neural networks (CNNs) face challenges in mobile and embedded vision applications due to high computational costs and memory requirements, which limit their effectiveness in resource-constrained devices and do not always require state-of-the-art accuracy for specific tasks.
Innovation Solution
The implementation of a lightweight CNN structure using separable convolutional layers with kernel sizes greater than 3×3 and residual shortcut connections, optimized for graphical processing units (GPUs), which reduces the number of convolutional blocks and layers needed to achieve desired accuracy and processing speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If convolutional neural networks increase in size and complexity to achieve higher accuracy, then accuracy is improved, but computational cost increases
Solution Approach 1:
The patent applies depthwise separable convolutions that divide the standard convolution operation into two independent parts: depthwise convolution (spatial filtering) and pointwise convolution (channel mixing). This segmentation reduces computational cost while maintaining accuracy by performing operations separately rather than simultaneously across all channels and spatial locations.
Solution Approach 2:
The patent changes the kernel size parameter to be greater than 3×3 (e.g., 5×5, 7×7) in the depthwise convolution layer. This parameter change increases the receptive field size, allowing the network to capture more spatial information with fewer computational operations, thereby improving accuracy without proportionally increasing computational cost.
2Measurement precision
If convolutional neural networks increase in size and complexity to achieve higher accuracy, then accuracy is improved, but memory requirements increase
Solution Approach 1:
By segmenting the convolution operation into depthwise and pointwise components, the patent reduces the number of parameters and computational operations required. The depthwise convolution uses larger kernels with fewer filters, while the pointwise convolution uses 1×1 kernels to combine channels, resulting in lower memory requirements compared to standard convolutions of equivalent accuracy.
3Measurement precision
If standard convolutional layers are used with larger kernel sizes, then accuracy is improved, but computational cost increases significantly
Solution Approach 1:
The patent segments the convolution into depthwise (spatial) and pointwise (channel) operations. The depthwise convolution with larger kernels processes each channel independently, reducing the multiplicative complexity, while the pointwise convolution then combines channels. This segmentation allows larger effective kernel sizes without the exponential computational cost increase of standard convolutions.
Solution Approach 2:
The patent changes the kernel size parameter to greater than 3×3 in the depthwise convolution while maintaining computational efficiency through the separable structure. This parameter change increases the receptive field without the full computational penalty of standard large-kernel convolutions because the depthwise and pointwise operations can be performed efficiently in sequence.
Data Source
AI summary
A computing system is disclosed including a convolutional neural configured to receive an input that describes a facial image and generate a facial object recognition output that describes one or more facial feature locations with respect to the facial image. The convolutional neural network can include a plurality of convolutional blocks. At least one of the convolutional blocks can include one or more separable convolutional layers configured to apply a depthwise convolution and a pointwise convolution during processing of an input to generate an output. The depthwise convolution can be applied with a kernel size that is greater than 3×3. At least one of the convolutional blocks can include a residual shortcut connection from its input to its output.