Contextual Multi-Scale CNN for Small Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional CNN-based object-detection methods struggle with detecting small, occluded, or low-resolution objects in images due to limited contextual information and inadequate feature representation, particularly in challenging conditions like heavy facial occlusions, extreme poses, and varying illumination.
Innovation Solution
The Contextual Multi-Scale Region-based CNN (CMS-RCNN) approach processes images by generating feature maps of multiple scales, incorporating contextual information from surrounding regions through Multiscale Region Proposal Network (MS-RPN) and Contextual Multi-Scale Region-based Convolutional Neural Network (CMS-CNN), which normalizes and concatenates feature maps to enhance object detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional CNN-based object-detection methods are used, then processing speed is maintained, but detection precision deteriorates for small, occluded, or low-resolution objects
Solution Approach 1:
The patent divides the feature extraction process into multiple independent convolutional layers that operate at different scales. Each layer processes features at a specific resolution level, allowing the system to capture both fine-grained details from small objects and contextual information from larger structures without requiring a single complex model
Solution Approach 2:
The patent introduces a multi-scale dimension by processing images at multiple resolutions simultaneously through different convolutional layers. This dimensional approach allows the system to analyze objects at various size levels, improving detection of small and occluded objects by examining them at appropriate scale levels rather than a single fixed resolution
2Reliability
If feature maps are processed without contextual information integration, then processing time is reduced, but detection reliability deteriorates in challenging conditions
Solution Approach 1:
The patent performs preliminary feature extraction at multiple scales before final object detection. By pre-processing images through multiple convolutional layers to generate scaled feature maps in advance, the system prepares contextual information that can be quickly referenced during detection, improving reliability without adding significant processing time during the actual detection phase
Solution Approach 2:
The patent merges feature maps from multiple convolutional layers that operate at different scales. By combining features from various levels of the convolutional network, the system integrates both local details and global context, improving detection reliability in challenging conditions while maintaining efficient processing through unified feature fusion
3Measurement precision
If single-scale feature maps are used, then computational complexity is reduced, but detection precision deteriorates for objects of varying sizes
Solution Approach 1:
The patent segments the feature processing into multiple convolutional layers, each responsible for extracting features at a specific scale. This segmentation allows the system to handle objects of varying sizes by assigning appropriate processing responsibilities to different layers, improving detection precision without requiring a single overly complex processing pipeline
Solution Approach 2:
The patent creates a multi-scale feature extraction system that serves multiple detection purposes simultaneously. The same convolutional network structure processes features at various scales, making the system universal for detecting objects of different sizes and resolutions without requiring separate specialized models for each object size category
Data Source
AI summary
Methods of detecting an object in an image using a convolutional neural-network-based architecture that processes multiple feature maps of differing scales from differing convolution layers within a convolutional network to create a regional-proposal bounding box. The bounding box is projected back to the feature maps of the individual convolution layers to obtain a set of regions of interest (ROIs) and a corresponding set of context regions that provide additional context for the ROIs. These ROIs and context regions are processed to create a confidence score representing a confidence that the object detected in the bounding box is the desired object. These processes allow the method to utilize deep features encoded in both the global and the local representation for object regions, allowing the method to robustly deal with challenges in the problem of object detection. Software for executing the disclosed methods within an object-detection system is also disclosed.