Vehicle Object Recognition Weak-Kernel Training With CNN
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional deep learning algorithms for autonomous driving focus on strengthening the model's current performance while neglecting its weaknesses, leading to reduced accuracy and increased uncertainty in complex environments.
Innovation Solution
A method and apparatus that generate misrecognition data to create a weak pattern map, classify a weak database, and selectively train a weak kernel of the object recognition model using a convolutional neural network (CNN) to address model weaknesses.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional deep learning algorithms train the model as a whole focusing on current performance, then the model strength is continuously learned, but the model weaknesses are further weakened and accuracy is reduced
Solution Approach 1:
The patent segments the model training process by identifying and isolating weak kernels from the overall model. Instead of training the entire model uniformly, the system extracts specific weak kernels that cause misrecognition and trains them separately using misrecognition data, thereby addressing the contradiction between maintaining overall strength and improving accuracy in weak areas
Solution Approach 2:
The patent applies local quality by focusing training efforts on specific local regions of the model (weak kernels) rather than the entire model. The system identifies which kernels are responsible for misrecognition and concentrates computational resources on improving only those specific kernels, allowing the model to maintain its overall strength while improving local accuracy
2Productivity
If the model is trained to resolve current problems, then the model strength is continuously learned, but the weaknesses of the model are further weakened
Solution Approach 1:
The patent implements feedback by using misrecognition data generated from comparing model predictions with ground truth to create a feedback loop. The weak pattern map is generated from misrecognition data, and the weak database is classified based on this map, creating a feedback mechanism that identifies and addresses model weaknesses systematically
Solution Approach 2:
The patent applies preliminary action by pre-generating misrecognition data and creating a weak pattern map before the actual training process. The weak database is classified in advance based on the weak pattern map, so that when training occurs, the system already has organized data and knowledge about which kernels need improvement, making the training process more efficient and targeted
Data Source
AI summary
A method and apparatus of processing an image and a vehicle including the same includes obtaining image information; inputting a ground truth; inputting the image information to an object recognition model to output recognition information including a position and a type of an object included in the image information; generating misrecognition data based on the output recognition information and the ground truth; generating a weak pattern map of the object recognition model by processing the misrecognition data; and classifying a weak database based on the weak pattern map and at least one piece of pre-stored image information.