Convolutional Neural Network Training with Dual Normalization Layers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current convolutional neural networks (CNNs) for image classification and detection primarily focus on texture features, neglecting shape features, which can lead to a trade-off between model robustness and accuracy, and may reduce the overall performance of the machine learning model.
Innovation Solution
The proposed method involves training a CNN using both original data and stylized data, where stylization reduces the emphasis on texture features while retaining shape features, employing separate normalization layers for each type of data to focus more on shape features, thereby enhancing robustness and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If adversarial examples are used in training to improve robustness, then model robustness is improved, but model accuracy deteriorates
Solution Approach 1:
The patent divides the training data into two distinct segments: original images and stylized versions of those images. This segmentation allows the model to learn different features from each segment - shape features from stylized images and texture features from original images - thereby resolving the contradiction between robustness and accuracy by treating them as separate learning objectives that can be balanced through dual normalization layers.
2Measurement precision
If training focuses on texture features, then texture recognition is improved, but shape feature learning deteriorates
Solution Approach 1:
The training data is segmented into original images for texture learning and stylized images for shape learning. The stylization process specifically preserves shape information while removing texture information, creating distinct training segments that target different feature types without interfering with each other.
Solution Approach 2:
Different normalization layers are applied to different data segments: a first normalization layer processes original images to preserve texture features, while a second normalization layer processes stylized images to preserve shape features. This local quality approach ensures each feature type is optimized independently according to its specific characteristics.
Data Source
AI summary
A computer-implemented training method for a convolutional neural network includes receiving first data and second data. The second data is data obtained after stylization is performed on the first data. The method further includes training the convolutional neural network based on the first data and the second data. The convolutional neural network has a first normalization layer and a second normalization layer. The first normalization layer is used for the first data, and the second normalization layer is used for the second data. The convolutional neural network trained in this way is no longer biased towards texture, and not only enhances robustness but also improves accuracy.