CNN Learning Apparatus for Edge Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Object detection CNNs require significant computational resources for high accuracy, making them challenging to implement on devices with limited resources, such as image sensors, and existing methods that improve accuracy often require additional computational power and skilled design.
Innovation Solution
A learning apparatus that trains a convolutional neural network (CNN) using ground truth labels and position information, allowing for parameter updates based on inferred value errors and position errors, enabling efficient training without separate updates for each error type, and emphasizing global or regional error information through weighting.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If object detection CNN is used to ensure high object recognition accuracy for any target scene, then object recognition accuracy is improved, but computational resources (high-performance computing unit, memory for holding half-computed processing) are significantly increased
Solution Approach 1:
The patent segments the computational task by separating feature extraction (performed by the image sensor's CNN) from object detection (performed by a separate processor). This allows the image sensor to focus only on extracting meaningful features with limited resources, while a more powerful processor handles the computationally intensive detection task using the extracted features.
Solution Approach 2:
The patent introduces an intermediary layer (feature extraction unit) between the image sensor and the object detection system. This intermediary processes the raw image data into compact feature representations, reducing the data volume and computational burden on subsequent detection stages while preserving essential information for accurate object recognition.
2Measurement precision
If object detection processing is added to achieve high accuracy for scenes where target object does not appear in entire image frame, then object recognition accuracy is improved, but it becomes difficult to implement under limited computational resources
Solution Approach 1:
The system divides functionality between the image sensor (feature extraction) and external processor (object detection), enabling the sensor to operate with limited resources while still achieving high accuracy detection capabilities through the collaborative architecture.
Solution Approach 2:
The image sensor is designed with multi-functionality, performing both traditional image capture and CNN-based feature extraction simultaneously. This universal approach allows the sensor to adapt to different application scenarios (full-frame and partial-frame objects) without requiring separate hardware configurations.
3Measurement precision
If additional processing (Bayesian network) is added to improve image classification accuracy, then classification accuracy is improved, but design complexity and computational resources are increased
Solution Approach 1:
The patent merges the classification functionality directly into the CNN architecture, combining feature extraction and classification into a unified neural network model. This integration eliminates the need for separate Bayesian network processing while maintaining high classification accuracy through the CNN's learned representations.
Solution Approach 2:
The CNN performs self-service by automatically learning optimal feature representations and classification boundaries from training data without requiring manual design of additional processing layers. The network adapts its parameters through backpropagation, eliminating the need for expert-designed Bayesian networks while achieving high accuracy.
Data Source
AI summary
A learning apparatus according to the present technology includes a learning unit that trains a CNN by using, as training data, a ground truth label prepared for each training image and ground-truth position information indicating a position of a target object in the training image.