Object Identification Using Segmented Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning techniques require large amounts of data and high computational resources for accurate image recognition, leading to increased memory capacity and processing time, as well as higher power consumption and costs.
Innovation Solution
A small-size neural network is trained using a reduced number of images, and the input image is modified through techniques like rotation, noise removal, brightness adjustment, and size adjustment to enhance object identification, with a system that includes a receiver, image modifier, and object determinator to improve recognition rates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning is used to ensure accurate object identification, then recognition accuracy is improved, but memory capacity and processing time are increased
Solution Approach 1:
The patent segments the object identification process into two distinct stages: a coarse identification stage using a small-size neural network for quick filtering, and a fine identification stage using a large-size neural network for detailed analysis. This segmentation allows the system to achieve accurate object identification while reducing overall memory capacity requirements, as the small network handles the majority of initial processing and only potentially triggers the larger network for ambiguous cases.
Solution Approach 2:
The patent applies partial action by using the small-size neural network to process all images initially, and only activating the large-size neural network when necessary (when confidence threshold is not met). This partial deployment of computational resources maintains high identification accuracy for critical cases while avoiding the excessive memory capacity and processing time that would result from always using the large network.
2Measurement precision
If deep learning is used to ensure accurate object identification, then recognition accuracy is improved, but processing time is increased
Solution Approach 1:
The patent segments the identification process into rapid coarse identification using a small neural network and detailed fine identification using a large neural network. The small network is trained quickly with less data and computational resources, providing fast initial results. This segmentation dramatically reduces the average processing time while maintaining accuracy through the two-stage approach.
Solution Approach 2:
The system performs partial deep learning processing by using the small-size neural network for the majority of cases and only invoking the time-consuming large-size neural network when necessary. This partial application of intensive processing maintains high accuracy for critical cases while avoiding excessive processing time for routine identifications.
3Use of energy by moving object
If a small size neural network is used to reduce memory capacity and processing time, then power consumption is reduced, but identification accuracy may deteriorate
Solution Approach 1:
The patent segments the identification system into a power-efficient small neural network for routine operations and a high-accuracy large neural network for critical cases. This segmentation enables the system to operate at low power consumption for the majority of cases while maintaining high identification accuracy when needed, effectively resolving the contradiction between power usage and precision.
Solution Approach 2:
The system applies partial deep learning by using the small neural network for most identifications and only activating the large neural network when the small network's confidence is insufficient. This partial use of high-power resources maintains identification accuracy while minimizing overall power consumption.
4Measurement precision
If image modification techniques are applied to enhance object identification, then recognition rate is improved, but processing complexity is increased
Solution Approach 1:
The patent applies preliminary action by performing image modifications (rotation, flipping, cropping, etc.) before feeding images to the neural network. These preprocessing steps prepare the data in advance, enhancing recognition rates by ensuring the network receives optimally formatted input. The complexity of these modifications is managed through automated preprocessing pipelines that integrate seamlessly with the two-stage identification system.
Data Source
AI summary
An artificial intelligence based object identifying apparatus and a method thereof which are capable of easily identifying a type of an object in an image using a small size learning model are disclosed. According to an embodiment of the present disclosure, an object identifying apparatus configured to identify an object from an image includes a receiver configured to receive the image, an image modifier configured to modify the received image by predetermined methods to generate a plurality of modified images, and an object determinator configured to apply the plurality of modified images to a neural network trained to identify an object from the image to obtain a plurality of identification results and determine a type of an object in the received image based on the plurality of identification results.