Dual-Submodel Training for High-Resolution Object Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current models for object recognition in autonomous driving require significant memory for internal calculations, limiting data processing capacity and efficiency, especially when handling high-resolution imaging data.
Innovation Solution
A method involving a first and second sub-model that down-scales and divides digital data into multiple areas, allowing for reduced memory requirements and computational effort through data scaling and division, enabling the model to process higher data quality or larger image sizes with the same memory capacity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high-resolution digital data is processed using traditional models, then object recognition accuracy is improved, but memory requirements increase significantly
Solution Approach 1:
The digital data is divided into multiple data areas, and the model is split into first and second sub-models. The first sub-model processes down-scaled data to capture global information, while the second sub-model processes individual data areas to capture local details. This segmentation allows the system to handle high-resolution data with reduced memory requirements by processing different regions separately rather than loading the entire high-resolution data at once.
Solution Approach 2:
The patent introduces a dimensional transformation by creating down-scaled versions of the digital data. The first sub-model operates on this reduced-dimensional data to extract global features, which are then combined with local features from the second sub-model. This dimensional change enables processing of high-resolution data within limited memory constraints.
2Quantity of substance
If data is divided into multiple areas and processed separately, then memory requirements are reduced, but processing complexity increases
Solution Approach 1:
The model is segmented into two specialized sub-models with distinct functions. The first sub-model handles down-scaled global data, while the second sub-model processes individual data areas. This segmentation distributes the processing complexity across specialized components rather than requiring a single complex model to handle all aspects simultaneously, making the overall system more manageable despite the increased number of components.
Solution Approach 2:
The outputs from the first and second sub-models are merged to produce the final result. The first output data (global information) and second output data (local details) are combined through addition or other operations. This merging strategy consolidates the complexity of processing multiple data areas into a unified output generation step, reducing the overall processing burden.
3Speed
If down-scaling is applied to digital data, then processing speed is improved, but data quality may be degraded
Solution Approach 1:
Different parts of the data are processed with different quality levels appropriate to their importance. The first sub-model processes down-scaled data at lower quality to capture global structures efficiently. The second sub-model processes individual data areas at higher quality to preserve local details. This local quality approach ensures that critical local features are maintained while allowing faster processing of less critical global information.
Solution Approach 2:
The results from low-quality down-scaled processing (first sub-model) and high-quality local processing (second sub-model) are merged. This combination restores the overall data quality by integrating global context from the down-scaled data with detailed local information from the individual data areas, achieving both speed and quality objectives.
Data Source
AI summary
A device and a method for training a model including a first sub-model and a second sub-model. Digital data are down-scaled to generate first input data. The digital data are divided into multiple data areas to generate second input data. A first sub-model generates first sub-model data relating to first input data fed to it. The first sub-model data are up-scaled to form first output data. A second sub-model for the data areas generates corresponding output data areas relating to second input data fed to it. The output data areas are assembled to form second output data. The first and second output data are combined to form third output data. The first sub-model is trained on the digital data by comparing provided target data and the first output data. The second sub-model is trained on the digital data by comparing the target data and the third output data.