Neural Network Proposal Selection for Efficient Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing computer vision applications and neural networks face challenges in accurately and efficiently detecting target images across large datasets due to intra-class variations and computational resource constraints, leading to significant processing time and resource consumption.
Innovation Solution
A system utilizing a neural network architecture with a region proposal network and a proposal selection model that generates a reduced proposal set by learning normal distributions, reducing the number of proposals to be analyzed, and using similarity scores to determine the presence of target objects, thereby minimizing processing time and resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional detection techniques are used to identify target images in large datasets, then detection accuracy can be maintained, but processing time and computational resources increase significantly
Solution Approach 1:
The patent segments the detection process into multiple stages: generating region proposals, selecting promising proposals using a proposal selection model, and performing detailed detection only on selected regions. This segmentation allows the system to avoid exhaustive analysis of all possible regions, thereby reducing processing time while maintaining detection accuracy through focused analysis of high-probability regions.
Solution Approach 2:
The patent employs preliminary action by generating region proposals and filtering them through a proposal selection model before performing the actual detection. This preliminary filtering step identifies and eliminates regions unlikely to contain target objects, allowing the main detection process to focus computational resources only on promising regions, thus reducing overall processing time without compromising accuracy.
2Reliability
If traditional detection techniques are used to identify target images in large datasets, then comprehensive detection can be achieved, but computational resources and energy consumption increase significantly
Solution Approach 1:
The detection process is segmented into proposal generation, proposal selection, and detection execution phases. By dividing the computational workload in this manner, the system achieves comprehensive detection coverage through the proposal selection model that identifies promising regions, while reducing overall computational resource consumption by avoiding exhaustive analysis of all possible regions.
Solution Approach 2:
The patent applies partial action by performing detailed detection only on a selected subset of region proposals rather than all possible regions. The proposal selection model identifies the most promising regions, and detection is performed partially on these selected regions, achieving sufficient comprehensiveness for practical applications while significantly reducing computational resource and energy consumption.
3Measurement precision
If the number of proposals to be analyzed is increased to improve detection accuracy, then more target objects can be detected, but processing time and computational load increase
Solution Approach 1:
The patent uses preliminary action through the proposal selection model that pre-filters region proposals before detailed detection. This preliminary step identifies and ranks proposals based on their likelihood of containing target objects, allowing the system to analyze a manageable number of high-probability regions while maintaining detection accuracy, thus preserving processing efficiency.
Solution Approach 2:
The patent changes the parameter of proposal quantity by dynamically selecting a reduced set of proposals based on confidence scores or ranking from the proposal selection model. Instead of analyzing all possible proposals, the system adjusts the number of proposals to be analyzed based on their predicted relevance, maintaining detection accuracy for high-probability cases while improving processing efficiency by limiting analysis to a manageable subset.
Data Source
AI summary
A systems including one or more processors and one or more non-transitory storage devices storing computing instructions configured to run on the one or more processors and perform acts of: receiving a query image comprising at least one target object and at least one candidate image; generating a plurality of proposals based on the at least one candidate image; selecting a portion of the plurality of proposals to produce a reduced proposal set; generating a query embedding corresponding to the query image; generating candidate embeddings corresponding to the portion of the plurality of proposals included in the reduced proposal set; computing similarity scores for the portion of the plurality of proposals; and comparing the similarity scores to a threshold to determine if the at least one candidate image comprises the at least one target object. Other embodiments are disclosed herein.