Shallow Neural Network Branches for Low-Power Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current object detection systems face challenges in achieving high accuracy at a low cost, particularly in real-time automotive applications where power consumption limits the use of deeper and wider neural networks, leading to ineffective detection of small objects due to insufficient semantic information and increased memory data transition complexities.
Innovation Solution
A method and system utilizing multiple branches with shallow neural networks and region units, where each branch is an instance of a single trained neural network, generates downscaled versions of input images to detect objects within a predefined size range, ignoring objects outside this range, thereby optimizing object detection without impairing speed or increasing power consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deeper and wider neural networks are used to improve object detection accuracy, then detection accuracy is improved, but power consumption increases and system speed decreases
Solution Approach 1:
The system segments the object detection task by creating multiple processing branches, each handling different scale ranges. This allows shallow networks to specialize in specific size ranges rather than requiring a single deep network to handle all scales, reducing overall power consumption while maintaining detection accuracy across varied object sizes.
Solution Approach 2:
The patent introduces a new dimension by processing images at multiple scales simultaneously through different branches. Instead of increasing network depth, the system adds scale diversity as a new dimension, allowing shallow networks to detect objects of different sizes by feeding them appropriately scaled versions of the input image.
2Measurement precision
If deeper neural networks are used to detect small objects with sufficient semantic information, then detection accuracy for small objects is improved, but device complexity increases
Solution Approach 1:
The system applies local quality by assigning different processing characteristics to different branches. Each branch is optimized for specific scale ranges, with shallow networks handling local scale-specific detection tasks. This localized specialization allows accurate small object detection without requiring globally complex deep network structures.
3Adaptability or versatility
If multiple scales are processed to detect objects of various sizes, then detection coverage is improved, but processing time increases
Solution Approach 1:
The patent merges multiple scale processing into parallel branches that operate simultaneously rather than sequentially. By combining the processing of different scales in parallel and using a selection unit to choose from candidate bounding boxes across branches, the system achieves multi-scale detection coverage without sequential processing delays.
Data Source
AI summary
A method that may include feeding an input image and downscaled versions of the input image to multiple branches of an object detector calculating, by the multiple branches, candidate bounding boxes; and selecting bounding boxes. The multiple branches comprise multiple shallow neural networks that are followed by multiple region units. Each branch includes a shallow neural network and a region unit. The multiple shallow neural networks are multiple instances of a single trained shallow neural network. The single trained shallow neural network is trained to detect objects having a size that is within a predefined size range and to ignore objects having a size that is outside the predefined size range.