Neural Network Object Detection Reducing Resource Consumption
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network systems for object detection in images are resource-intensive, making them unsuitable for implementation on devices with limited computing resources such as mobile phones and tablets.
Innovation Solution
A neural network system that uses initial convolutional neural network layers to generate feature maps, followed by location and confidence score generating convolutional neural network layers, which perform specific convolutions to predict object locations and confidence scores, reducing the need for fully-connected layers and thus minimizing computing resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional neural network systems with fully-connected layers are used for object detection, then detection accuracy can be achieved, but computing resource consumption increases significantly
Solution Approach 1:
The patent segments the neural network into distinct functional modules: initial convolutional layers for feature extraction, location generating convolutional layers for bounding box prediction, and confidence score generating convolutional layers for detection confidence. This modular segmentation replaces resource-intensive fully-connected layers with efficient convolutional operations, reducing computing resource consumption while preserving detection accuracy through specialized function assignment to each segment
Solution Approach 2:
The patent transitions from traditional 2D convolutional operations to 3D convolutions by incorporating the time dimension in video processing. The location and confidence score generating layers operate across multiple frames simultaneously, leveraging temporal information to improve detection accuracy while maintaining computational efficiency through shared convolutional kernels across the temporal dimension
2Measurement precision
If resource-intensive neural network architectures are deployed, then better object detection performance is achieved, but device compatibility is reduced
Solution Approach 1:
The patent substitutes traditional mechanical neural network architectures (fully-connected layers) with convolutional neural network operations that are natively supported by mobile processors and GPUs. This substitution enables the system to run on resource-constrained devices like mobile phones and tablets while maintaining detection performance, as convolutional operations are more computationally efficient and better optimized for mobile hardware
3Productivity
If fully-connected layers are used for location and confidence score generation, then comprehensive feature processing is achieved, but computational complexity increases
Solution Approach 1:
The patent merges the location generation and confidence score generation functions into unified convolutional neural network layers that operate simultaneously on the same feature maps. The location generating convolutional layers and confidence score generating convolutional layers process features in parallel, reducing architectural complexity compared to separate fully-connected layers while maintaining comprehensive feature processing capability
Data Source
AI summary
Systems, methods, and apparatus, including computer programs encoded on a computer storage medium. In one aspect, a system includes initial neural network layers configured to: receive an input image, and process the input image to generate a plurality of first feature maps that characterize the input image; a location generating convolutional neural network layer configured to perform a convolution on the representation of the first plurality of feature maps to generate data defining a respective location of each of a predetermined number of bounding boxes in the input image, wherein each bounding box identifies a respective first region of the input image; and a confidence score generating convolutional neural network layer configured to perform a convolution on the representation of the first plurality of feature maps to generate a confidence score for each of the predetermined number of bounding boxes in the input image.