SAR Image Feature Enhancement via Transfer Learning and Auto-Encoder Mapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge in the Guangdong-Hong Kong-Macao Greater Bay Area is the insufficient labeled sample size for SAR images, leading to over-fitting, slow convergence, and degraded generalization capability in SAR target detection, particularly in complex scenes, which hampers accurate feature extraction and recognition.
Innovation Solution
A method and apparatus that employ transfer learning and network optimization using deep neural networks, specifically initializing networks with virtual images and utilizing auto-encoders to enhance semantic features, improving convergence speed and recognition accuracy by mapping feature maps between feature and semantic spaces.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional feature extraction models are trained with insufficient labeled SAR image samples, then training can be performed, but over-fitting occurs and generalization capability degrades
Solution Approach 1:
The patent applies data augmentation by generating virtual SAR images through copying and transforming existing real SAR images. Virtual images are created by applying geometric transformations, noise addition, and other processing to real images, thereby expanding the training dataset without requiring additional physical imaging resources. This resolves the contradiction by artificially increasing sample quantity while maintaining data quality and generalization capability.
Solution Approach 2:
The patent implements pre-training on large-scale natural images (ImageNet) before fine-tuning on SAR images. This preliminary action allows the model to learn general feature representations from abundant natural image data, then adapt these representations to SAR-specific features. This two-stage approach enables effective training with limited SAR samples by leveraging knowledge transferred from the pre-training phase.
2Speed
If conventional models are trained with small sample sizes, then training can proceed, but convergence speed becomes slow
Solution Approach 1:
The patent performs pre-training on large-scale natural images before fine-tuning on SAR images. This preliminary training establishes good initial weights and feature representations, enabling the model to converge much faster when trained on the limited SAR image samples. The pre-trained model serves as an excellent starting point, reducing the number of iterations needed to achieve optimal performance.
Solution Approach 2:
By generating virtual SAR images through data augmentation, the patent effectively increases the training dataset size. This enlarged dataset provides more training examples per epoch, allowing the model to learn more robust features and converge faster despite the originally limited number of real SAR images.
3Measurement precision
If deep neural networks are trained with insufficient samples, then feature extraction can be performed, but over-fitting occurs
Solution Approach 1:
The patent generates virtual SAR images by copying and transforming real SAR images through various augmentations including geometric transformations, noise addition, and intensity modifications. This creates a diverse training dataset that prevents the model from memorizing specific samples, thereby reducing over-fitting while maintaining high feature extraction accuracy.
Solution Approach 2:
The patent implements pre-training on large-scale natural images to establish robust feature extraction capabilities before adapting to SAR images. This preliminary training on diverse natural images provides strong regularization effects, preventing over-fitting when the model is subsequently trained on limited SAR images while maintaining high measurement precision.
4Speed
If transfer learning is applied to initialize deep neural networks, then convergence speed increases, but model complexity increases
Solution Approach 1:
The patent implements a two-stage training process where the model is first pre-trained on ImageNet and then fine-tuned on SAR images. While this adds an initialization step, the process is automated and standardized, making the increased complexity manageable. The benefits of faster convergence and better performance outweigh the additional initialization complexity.
Data Source
AI summary
The present disclosure relates to a method for enhancing sematic features of SAR image oriented small set of samples, comprising: acquiring a sample set of an SAR target image, and performing transfer learning and training on the sample set to obtain a initialized deep neural network of an SAR target image, the sample set comprising an SAR target image and an SAR target virtual image; performing network optimization on the deep neural network by an activation function, and extracting features of the SAR target image by the optimized deep neural network to obtain a feature map; and mapping, by an auto-encoder, the feature map between a feature space and a semantic space to obtain a deep visual feature with an enhanced semantic feature.