Residual CNN Object Detection for Microcontroller Edge Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing CNN-based object detection methods require significant computational power and memory, making them unsuitable for edge devices like microcontrollers due to the lack of accelerators.
Innovation Solution
A lightweight convolutional neural network system utilizing depth-wise separable convolutions and residual structures to reduce computational and memory requirements, enabling efficient object detection on microcontrollers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If standard CNN-based object detection methods (YOLO, SSD, RetinaNet) are used, then detection accuracy is improved, but computational power and memory requirements increase significantly
Solution Approach 1:
The patent applies segmentation by dividing the convolution operation into two separate operations: standard convolution and depth-wise separable convolution. This segmentation allows the network to process features in stages, reducing the computational complexity from O(k^3) to O(k^2) where k is the kernel size, thereby lowering power requirements while maintaining detection accuracy
Solution Approach 2:
The patent changes the parameter configuration by using different convolution scales (first scale and second scale) in different network stages. This parameter variation allows the model to adapt computational intensity to different feature extraction needs, optimizing the balance between accuracy and power consumption
2Measurement precision
If standard CNN-based object detection methods are used, then detection capability is improved, but memory requirements increase significantly
Solution Approach 1:
The patent segments the feature extraction process into multiple stages with different convolution operations. By processing features in stages rather than all at once, the network reduces the peak memory footprint while maintaining the ability to detect objects accurately
Solution Approach 2:
The patent extracts and processes only the most relevant features at each stage using selective convolution operations. This extraction approach avoids processing redundant information, reducing memory requirements while preserving detection capability
3Power
If lightweight convolution operations are used, then computational requirements are reduced, but detection precision may deteriorate
Solution Approach 1:
The patent merges the results of standard convolution and depth-wise separable convolution through addition operations. This merging combines the strengths of both operations: the feature extraction capability of standard convolution and the computational efficiency of depth-wise separable convolution, achieving both reduced computational requirements and maintained precision
Solution Approach 2:
The patent creates a composite convolutional network structure that combines different convolution types (standard and depth-wise separable). This composite approach allows the network to leverage multiple processing strategies, maintaining detection precision while reducing overall computational burden
4Measurement precision
If multiple convolution operations with different scales are applied, then feature extraction capability is improved, but device complexity increases
Solution Approach 1:
The patent segments the network into distinct blocks (initial extraction block, residual blocks, extraction output block) with specific functions. This segmentation makes the complex multi-scale convolution process more manageable and implementable on resource-constrained devices by organizing operations into reusable modular units
Data Source
AI summary
A method for detecting a target object using a convolutional neural network includes: receiving an image frame; producing a first feature map of the image frame; producing a residual feature map from the first feature map by: applying a convolution of a first scale, and thereafter a depth-wise separable convolution, to the first feature map, thereby producing a second feature map; and adding the first feature map and the second feature map to produce an added feature map; producing at least one extracted feature map of the image frame, from the residual feature map by applying additional convolution of a second scale different from the first scale, and thereafter additional depth-wise separable convolution, to the residual feature map; and determining a box using the at least one extracted feature map. The determined box frames the target object.