Convolutional Neural Network Architecture for Edge Object Detection

Solution Overview

Problem

Existing object detection systems based on neural networks are inefficient on edge devices due to slow operation speeds, which fail to meet the requirements of various application programs.

Innovation Solution

An object detection system and method utilizing a convolutional neural network architecture with multiple scales, where the network width of certain layers is adjusted to increase operation speed while maintaining detection accuracy, by reducing the network width of secondary convolution layers and maintaining or increasing the network width of initial layers.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If the network width of convolution layers is increased to improve detection accuracy, then the operation speed decreases and processing time increases

Engineering Contradiction:

Improvedetection accuracyVSAvoidoperation speed

Core Design Contradiction:

Measurement precisionVSSpeed

Solution Approach 1:

The patent applies different network widths to different convolution layers based on their specific functions. The first convolution layer maintains a larger network width (e.g., 512 channels) to preserve feature extraction quality for detection accuracy, while the second convolution layer uses a reduced network width (e.g., 128 channels) to accelerate processing. This localized optimization resolves the contradiction by making each layer's complexity match its functional requirements rather than uniformly increasing all layers.

Inventive Principle:

Principle #3Local quality

Solution Approach 2:

The convolutional neural network is segmented into multiple scales with different network widths. The first scale includes convolution layers with larger network widths for detailed feature extraction, while the second scale includes convolution layers with reduced network widths for faster processing. This segmentation allows the system to achieve both high detection accuracy in critical layers and fast operation speed in less critical layers.

Inventive Principle:

Principle #1Segmentation

2Measurement precision

If the network width of convolution layers is increased to improve detection accuracy, then the processing time increases

Engineering Contradiction:

Improvedetection accuracyVSAvoidprocessing time

Core Design Contradiction:

Measurement precisionVSLoss of time

Solution Approach 1:

The patent applies different network widths to different convolution layers based on their specific functions. The first convolution layer maintains a larger network width (e.g., 512 channels) to preserve feature extraction quality for detection accuracy, while the second convolution layer uses a reduced network width (e.g., 128 channels) to accelerate processing. This localized optimization resolves the contradiction by making each layer's complexity match its functional requirements rather than uniformly increasing all layers.

Inventive Principle:

Principle #3Local quality

Solution Approach 2:

The convolutional neural network is segmented into multiple scales with different network widths. The first scale includes convolution layers with larger network widths for detailed feature extraction, while the second scale includes convolution layers with reduced network widths for faster processing. This segmentation allows the system to achieve both high detection accuracy in critical layers and fast operation speed in less critical layers.

Inventive Principle:

Principle #1Segmentation

3Measurement precision

If a complete convolutional neural network architecture is used to maintain detection accuracy, then the device complexity increases

Engineering Contradiction:

Improvedetection accuracyVSAvoidnetwork architecture complexity

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent applies different network widths to different convolution layers based on their specific functions. The first convolution layer maintains a larger network width (e.g., 512 channels) to preserve feature extraction quality for detection accuracy, while the second convolution layer uses a reduced network width (e.g., 128 channels) to accelerate processing. This localized optimization resolves the contradiction by making each layer's complexity match its functional requirements rather than uniformly increasing all layers.

Inventive Principle:

Principle #3Local quality

Solution Approach 2:

The network architecture dynamically adjusts the number of channels in different convolution layers. The first convolution layer uses more channels (e.g., 512) for comprehensive feature extraction, while the second convolution layer uses fewer channels (e.g., 128) to reduce computational complexity. This dynamic configuration allows the system to maintain detection accuracy where needed while reducing overall device complexity.

Inventive Principle:

Principle #15Dynamics

Data Source

PatentUS11195083B2Object detection system and object detection method

Publication Date: 2021.12.07 PEGATRON

AI summary

An object detection system includes an image capture device, a memory, and a processor. The image capture device captures an image. The memory stores an instruction corresponding to an inference engine based on a multi-scale convolutional neural network architecture including a first, a second, and an object detection scale. The processor executes the instruction to: reduce network widths of convolution layers of the second scale; run the inference engine according to the adjusted convolutional neural network architecture to receive the image as an initial input; input a first output generated by the first scale according to the initial input into the second and the object detection scale; input a second output generated by the second scale according to the first output into the object detection scale; generate a final output according to the first and the second output by the object detection scale, to perform object detection on the image.