Multiscale Point Cloud Classification via Parallel Network Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing point cloud classification methods rely heavily on manual feature extraction, which is complex and limited to specific tasks, and deep learning methods often ignore local relationships between points, leading to low accuracy and robustness, as well as lengthy training times.
Innovation Solution
A multiscale point cloud classification method using a pre-trained parallel classification network that includes multiple basic networks with feature alignment, extraction, and fully connected layers, combined with a deep Q network for feature fusion to improve classification accuracy and robustness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If deep learning methods are used to extract global features of point cloud, then automation degree is improved, but local relationships between points are ignored leading to low accuracy and robustness
Solution Approach 1:
The patent divides the point cloud processing into multiple scales and levels, using a hierarchical approach where global features and local features are extracted separately and then integrated. This segmentation allows both global context and local relationships to be captured, resolving the contradiction between automation and accuracy.
Solution Approach 2:
The patent introduces a multi-scale dimensional framework that processes point clouds at different resolution levels. By adding the scale dimension, the system can simultaneously consider global structures (at coarser scales) and local relationships (at finer scales), thereby improving classification accuracy while maintaining automation.
2Extent of automation
If deep learning methods are used for point cloud classification, then automation degree is improved, but training time increases significantly
Solution Approach 1:
The patent employs pre-trained networks and feature extractors that have been trained in advance on large datasets. This preliminary action allows the system to achieve high automation with reduced training time for specific applications, as the heavy lifting of learning fundamental features has already been done during pre-training.
Solution Approach 2:
The training process is segmented into distinct phases: pre-training on large datasets, then fine-tuning on specific tasks. This segmentation allows the system to benefit from automated feature learning while controlling overall training time by reusing pre-learned representations rather than training from scratch.
3Loss of time
If manual feature extraction is used for point cloud classification, then training time is reduced, but the process becomes complex and limited to specific tasks
Solution Approach 1:
The patent replaces manual mechanical feature extraction processes with automated deep learning-based feature extraction. This substitution reduces the need for expert intervention and simplifies the overall process while maintaining flexibility across different tasks, thereby reducing both time and complexity.
Solution Approach 2:
The patent employs universal feature extractors and pre-trained networks that can be applied across multiple tasks and datasets. This multi-functionality eliminates the need for task-specific manual feature engineering, reducing process complexity while enabling automated classification across diverse applications.
4Device complexity
If only global features are extracted from point cloud, then computation is simplified, but classification accuracy and robustness decrease
Solution Approach 1:
The patent segments feature extraction into global and local components, computing each separately and then integrating them. This segmentation allows the system to maintain computational efficiency for global features while adding local feature computation only where needed, thereby improving accuracy without excessively increasing complexity.
Solution Approach 2:
The patent applies local quality analysis selectively at different scales, focusing computational resources on extracting local relationships where they provide the most value. This approach maintains overall computation simplicity while enhancing classification accuracy through targeted local feature extraction.
Data Source
AI summary
The present disclosure discloses a multiscale point cloud classification method. The method includes the following steps: acquiring 3D unordered point cloud data; performing feature extraction and classification on the acquired point cloud data using a pre-trained parallel classification network to obtain an output result, wherein the parallel classification network includes a plurality of basic networks with the same structures; and fusing the output results of the parallel network using a pre-trained deep Q network to obtain a final result of point cloud classification. The present disclosure can improve the accuracy and robustness of point cloud classification.