Recursive Cortical Network for Object Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current artificial intelligence systems, such as convolution neural networks and Hierarchical Temporal Memory, fail to achieve ideal recognition performance in object recognition and generation tasks, particularly in dealing with variations in object positions, settings, and types, and struggle with detecting multiple objects in a scene.
Innovation Solution
A recursive cortical network (RCN) is introduced, comprising sub-networks with parent feature nodes, pool nodes, parent-specific child feature nodes, and constraint nodes, which implement summation, multiplication, and other functions to enhance invariance, selectivity, and information sharing, allowing for improved pattern inference and generation across hierarchical layers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing object recognition models (convolution neural networks, HMAX, SFA, HTM) are used, then basic pattern recognition is achieved, but recognition performance falls short of ideal levels particularly in handling object variations and multiple objects
Solution Approach 1:
The system segments the image into multiple candidate regions using detection algorithms, then applies specialized recognition networks to each region. This segmentation approach allows the system to handle multiple objects simultaneously and focus computational resources on relevant areas, improving both accuracy and adaptability to various object scenarios.
Solution Approach 2:
The system transitions from standard 2D image processing to 3D volumetric processing by constructing three-dimensional representations from multiple 2D slices. This dimensional transformation enables better handling of object variations by providing depth information and spatial context that enhances recognition accuracy for objects in different positions and orientations.
2Measurement precision
If standard 2D image processing is used, then processing speed is maintained, but the system fails to capture spatial relationships and depth information necessary for ideal recognition
Solution Approach 1:
The system constructs 3D volumetric representations by stacking and processing multiple 2D image slices through specialized neural networks. This approach captures spatial relationships and depth information by transforming the problem into three-dimensional space, enabling the system to understand object geometry and spatial context while maintaining computational feasibility through efficient network architectures.
Solution Approach 2:
The processing architecture is segmented into distinct functional components: detection modules that identify candidate regions, recognition networks that process each region, and integration systems that combine results. This segmentation manages complexity by dividing the overall task into manageable sub-tasks, each handled by specialized network components.
3Measurement precision
If computational resources are increased to improve recognition accuracy, then pattern detection performance improves, but the system becomes less efficient and more resource-intensive
Solution Approach 1:
The system divides the image processing task into segments, applying computational resources selectively to candidate regions rather than processing the entire image uniformly. This region-based approach improves efficiency by focusing computation on areas containing objects of interest while reducing overall computational burden compared to full-image processing.
Solution Approach 2:
The system applies recognition processing to a subset of candidate regions identified by detection algorithms rather than processing all possible regions. This partial action approach balances accuracy and efficiency by applying intensive computation only where necessary, avoiding wasteful processing of areas that do not contain objects of interest.
Data Source
AI summary
A method for inferring patterns in multi-dimensional image data comprises providing a recursive network of sub-networks with a parent feature node and at least two child feature nodes; wherein each sub-network is associated with a distinct subset of the space; configuring nodes of the sub-networks with posterior distribution component; receiving image data feature input at the final child feature nodes; propagating node activation through the network layer hierarchy in a manner consistent with node connections of sub-networks of the network and the posterior prediction of child nodes; and outputting parent feature node selection to an inferred output.