Dual Neural Network Robot Object Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep neural networks used for image classification in traffic scenes often misclassify objects due to varying lighting conditions and inadequate training data, leading to incorrect object detection and classification.
Innovation Solution
A method utilizing two machine learning systems, where a first convolutional neural network provides initial classifications, and a second system, comprising a recurrent or fully connected neural network, integrates temporal information to enhance classification precision by reassessing previous classifications and determining a confidence level for accurate object characterization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a single deep neural network is used for object classification, then the system structure remains simple, but classification accuracy decreases under varying lighting conditions and perspectives
Solution Approach 1:
The patent divides the classification task into multiple sequential stages using different neural network architectures. The first CNN processes individual images for initial classification, while the second RNN processes sequences of classifications over time. This segmentation allows each component to specialize in specific aspects of the classification problem, improving overall accuracy without requiring a single overly complex network.
Solution Approach 2:
The patent transitions from spatial processing (single image classification by CNN) to temporal processing (sequence of classifications by RNN). By adding the time dimension and processing sequences of classifications rather than individual images, the system captures temporal patterns and contextual information that improve accuracy under varying conditions.
2Reliability
If training data is limited or incomplete, then data collection requirements are reduced, but object detection reliability decreases
Solution Approach 1:
The RNN processes sequences of classifications and uses temporal context to correct individual misclassifications. By providing feedback from previous and subsequent classifications in the sequence, the system can identify and correct errors that would be missed by single-image classification, thereby improving reliability without requiring more training data.
Solution Approach 2:
The system performs preliminary classification with the first CNN, then uses the second RNN to reassess and refine these classifications over time. This preliminary action followed by temporal refinement allows the system to achieve higher reliability by catching errors that would otherwise require extensive training data to prevent.
3Measurement precision
If temporal information from image sequences is utilized, then classification precision improves, but processing time increases
Solution Approach 1:
The patent segments the processing into two distinct phases: rapid initial classification by the first CNN on individual images, followed by temporal refinement by the second RNN on sequences. This segmentation allows the system to quickly process individual frames while only applying the more computationally intensive temporal analysis where needed, balancing precision and processing time.
4Reliability
If multiple machine learning systems are deployed, then misclassification detection capability improves, but computational resources consumption increases
Solution Approach 1:
The patent implements partial redundancy by using two neural network systems, but the second RNN processes only the output sequences from the first CNN rather than reprocessing the original images. This partial action approach provides misclassification detection capability while consuming fewer computational resources than would be required for complete redundant processing of raw data.
Data Source
AI summary
A method for operating e.g. a robot. The method includes the steps: acquiring a sequence of images of a space surrounding the robot within which a potential object is situated; ascertaining, using a first machine learning system, a respective first variable characterizing the potential object as a function of each of the images; ascertaining, using a second machine learning system, at least one second variable that characterizes the potential object as a function of a plurality of the first variables characterizing the potential object; controlling the robot as a function of the second variable characterizing the potential object. A computer program, a device for carrying out the method, and a machine-readable storage element on which the computer program is stored, are also described.