Image Segmentation Network for Small Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep convolutional neural networks (CNNs) face challenges in detecting small objects and have memory limitations due to the size of input data, which restricts their effectiveness in applications like autonomous driving.
Innovation Solution
A deep neural network pipeline that includes multiple convolutional and deconvolutional layers, along with filter kernel interpolation, to enhance object detection in camera-captured images, allowing for better initialization and faster convergence, and reducing memory requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the input size is increased to improve small object detection, then detection accuracy improves, but memory requirements increase
Solution Approach 1:
The patent divides the image processing into multiple stages with different input sizes. Stage 1 processes the full-resolution image to detect small objects, while Stage 2 processes a downsampled version for large objects. This segmentation allows the system to maintain high detection accuracy for small objects without requiring the entire network to handle unnecessarily large input sizes, thus optimizing memory usage.
Solution Approach 2:
The patent implements dynamic input size selection where the network adaptively processes different regions of the image at different resolutions. Small objects in the image are processed at full resolution while larger regions are processed at lower resolutions, allowing the system to optimize memory requirements while maintaining detection accuracy for small objects.
2Measurement precision
If the network depth is increased to improve feature extraction, then object classification accuracy improves, but computational complexity increases
Solution Approach 1:
The patent segments the deep network into two distinct stages with different depths and functionalities. Stage 1 uses a shallower network (e.g., VGG16) focused on detecting small objects, while Stage 2 uses a deeper network (e.g., VGG19) for comprehensive feature extraction and classification. This segmentation allows each stage to be optimized for its specific purpose, improving overall classification accuracy without unnecessarily increasing the complexity of every part of the network.
Solution Approach 2:
The patent creates a multi-functional network architecture where Stage 1 serves the specific function of small object detection while Stage 2 provides general-purpose feature extraction and classification. This multi-functionality allows the system to achieve high classification accuracy across different object sizes without requiring a single overly complex network to handle all scenarios.
3Measurement precision
If the input image resolution is increased to detect small objects, then detection precision improves, but processing time increases
Solution Approach 1:
The patent segments the processing pipeline to handle different object sizes efficiently. Stage 1 processes the full-resolution image specifically for small object detection, ensuring high precision for small objects. Stage 2 processes a downsampled version of the image for larger objects, reducing processing time for regions where high resolution is not necessary. This segmentation optimizes the trade-off between detection precision and processing time.
Solution Approach 2:
The patent applies partial high-resolution processing only to regions where small objects are likely to be present, rather than processing the entire image at full resolution throughout the network. This partial action maintains detection precision for small objects while reducing overall processing time by avoiding unnecessary high-resolution processing in regions where it is not needed.
Data Source
AI summary
An apparatus and method for encoding objects in a camera-captured image with a deep neural network pipeline including multiple convolutional neural networks or convolutional layers. After identifying at least a portion of the camera-capture image, a first convolutional layer is applied to the at least the portion of the camera-captured image and multiple subregion representations are pooled from the output of the first convolutional layer. One or more additional convolutions are performed. At least one deconvolution is performed and concatenated with the output of one or more convolutions. One or more final convolutions are performed. The at least the portion of the camera-captured image is classified as an object category in response to an output of the one or more final convolutions.