Neural Sensor Training for Bulk Flow Classification With Low-Cost Imaging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning systems for classifying objects in a bulk flow are inefficient and economically unfeasible due to the need for high-quality sensor data, which requires expensive sensors, limiting their applicability in sorting applications.
Innovation Solution
A method of training a neural network using input and auxiliary image data from sensors of different or varying technologies to enhance classification accuracy and efficiency, allowing for cost-effective classification and sorting of objects in a bulk flow.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If expensive sensors are used to capture high-quality sensor data for machine learning classification, then classification accuracy is improved, but system cost increases
Solution Approach 1:
The patent creates virtual copies of expensive sensor data by using synthetic image data generated through rendering engines. These synthetic copies mimic real sensor outputs but are produced at minimal cost, allowing the neural network to be trained without requiring actual expensive sensors during the training phase.
Solution Approach 2:
The patent replaces expensive, durable sensors with inexpensive synthetic data generation processes. The synthetic image data serves as a disposable alternative that can be generated on-demand without the need for physical sensor hardware, significantly reducing system costs while maintaining training effectiveness.
2Measurement precision
If a machine learning system processes a single stream of objects at a time, then classification accuracy is maintained, but processing speed decreases
Solution Approach 1:
The patent performs preliminary actions by pre-processing and augmenting synthetic training data before actual classification tasks. The neural network is pre-trained on extensively augmented synthetic datasets that simulate various bulk flow conditions, enabling it to quickly and accurately classify objects in bulk streams without requiring slow sequential processing during operation.
3Ease of manufacture
If synthetic image data is used to train the neural network, then system cost is reduced, but training data quality may be compromised
Solution Approach 1:
The patent merges multiple synthetic image datasets with different augmentations, rendering conditions, and simulated sensor characteristics into a comprehensive training corpus. This combination creates a rich, diverse training dataset that compensates for the synthetic nature of individual data points, maintaining high training data quality while using cost-effective synthetic generation methods.
Solution Approach 2:
The patent systematically varies parameters in synthetic image generation including lighting conditions, object positions, camera angles, and sensor noise characteristics. These parameter changes create diverse training scenarios that improve the robustness and generalization capability of the neural network, ensuring high training data quality despite using synthetic sources.
Data Source
AI summary
A method of training a neural network stored on a computer-readable storage medium to classify objects in a bulk flow, the method including: providing input image data depicting objects to be classified, which input image data is captured by means of an input imaging sensor of a first sensor technology design; providing auxiliary image data, which auxiliary image data is captured by means of an auxiliary imaging sensor of a second sensor technology design, and which auxiliary image data depicts said or similar objects which are classified in accordance with a predetermined classifying scheme; by means of a processing unit, train the neural network stored on the computer-readable storage medium to classify the depicted objects in the input image data based on classifications of depicted objects in the auxiliary image data, wherein the depicted objects in the input image data correspond to objects in a bulk flow.