Deep Learning Network Pruning for Real-Time Object Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current automated systems for analyzing image data struggle to provide real-time or near real-time object classification and identification with high accuracy, especially in vast datasets, as they are not optimized to track vector-based information such as spatial relationships between objects, leading to inefficiencies and inaccuracies in inference processes.
Innovation Solution
The implementation of a deep learning network with optimized vector concatenation (OSVC) is used within an imaging system, which includes reconfigurable focal plane electronics (RFPE) and GPUs, to propagate pose information and selectively prune unnecessary network branches, enhancing runtime performance and accuracy by focusing on significant vector information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a deep learning network processes all image data to improve object classification accuracy, then measurement precision improves, but processing time increases
Solution Approach 1:
The patent segments the deep learning network into multiple processing stages: a first neural network performs initial object detection and generates candidate regions, while a second neural network focuses processing only on these candidate regions for final classification. This segmentation allows the system to maintain high accuracy while reducing overall processing time by avoiding full-network processing of entire images.
Solution Approach 2:
The patent extracts and processes only the most relevant portions of image data. The first neural network identifies candidate regions containing potential objects, and the second neural network extracts and processes only these candidate regions rather than the entire image. This extraction approach maintains classification accuracy while significantly reducing processing time and computational resources required.
2Measurement precision
If vector information is propagated through the deep learning network to improve object identification accuracy, then measurement precision improves, but device complexity increases
Solution Approach 1:
The patent applies local quality by processing different types of information through specialized network components. The first neural network focuses on detecting candidate regions with basic features, while the second neural network focuses on precise classification of specific objects within those regions. This localized specialization allows vector information to be propagated effectively without requiring the entire network to handle all complexity.
Solution Approach 2:
The patent transitions from two-dimensional image processing to three-dimensional pose estimation by propagating vector information through the network. The system not only identifies objects but also determines their spatial orientation, position, and configuration in 3D space. This dimensional enhancement improves identification accuracy while the multi-stage architecture manages the associated complexity.
3Productivity
If the deep learning network is optimized by pruning unnecessary branches to improve processing speed, then productivity improves, but measurement precision may deteriorate
Solution Approach 1:
The patent applies preliminary action by using the first neural network to pre-process images and identify candidate regions before the second neural network performs detailed classification. This preliminary filtering step reduces the amount of data that requires intensive processing, improving overall speed while maintaining accuracy by ensuring that the second network focuses only on relevant regions that require precise classification.
Data Source
AI summary
Imaging systems and methods that implement a deep learning network are disclosed. The deep learning network utilizes pose information associated with at least some identified objects. The network is pruned, to reduce the amount of information processed and to optimize runtime processing when the network is deployed. In operation, the network identifies objects, and propagates pose information for at least some of the objects or components of identified objects. The network can be deployed as part of a processing system of an imaging system included as part of a remote platform.