Lightweight Semantic Segmentation Training With Dual Teacher Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for training lightweight semantic segmentation models on low-computing-resource devices result in significant performance degradation, affecting image segmentation capability, generalization, and stability due to the use of traditional teacher models.
Innovation Solution
A semantic segmentation model training method utilizing a pre-trained teacher network with low depth and high width, and another with high depth and low width, to generate complementary segmentation maps for training a lightweight student model, incorporating supervised and unsupervised losses to optimize the student model's performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional teacher models are used for knowledge distillation, then the student model can be trained, but the student model suffers from significant performance degradation in image segmentation capability, generalization, and stability
Solution Approach 1:
The patent segments the teacher model into two specialized networks: a shallow teacher network (fewer layers, better local feature extraction) and a deep teacher network (more layers, better global semantic understanding). Each teacher network processes the input image independently to generate separate segmentation maps, which are then fused to provide comprehensive supervision for the student model, preventing information loss and performance degradation.
Solution Approach 2:
The patent introduces a new dimension in model architecture by varying network depth while maintaining comparable parameter scales. The shallow teacher network focuses on local details with fewer layers, while the deep teacher network captures global semantics with more layers. This dimensional differentiation allows the student model to learn both local and global features effectively, improving segmentation capability without the performance degradation caused by traditional single-teacher approaches.
2Device complexity
If a lightweight student model is trained for deployment on low-computing-resource devices, then deployment feasibility is improved, but model performance and generalization capability deteriorate
Solution Approach 1:
The patent applies local quality by assigning different architectural characteristics to different teacher networks while keeping the student model lightweight. The shallow teacher network uses fewer layers optimized for local feature extraction, while the deep teacher network uses more layers for global semantic understanding. The student model maintains lightweight structure but receives differentiated supervision signals from both teachers, enabling it to achieve good generalization capability without increasing device complexity.
3Ease of operation
If knowledge distillation is performed using a single teacher model, then the training process is simple, but the student model fails to learn both local details and global semantics effectively
Solution Approach 1:
The training process is segmented into two parallel knowledge distillation streams: one from the shallow teacher network focusing on local details and another from the deep teacher network focusing on global semantics. Both streams operate independently but contribute to the same student model training objective. This segmentation maintains relative training simplicity while dramatically improving feature extraction accuracy by providing comprehensive supervision from both local and global perspectives.
Data Source
AI summary
A semantic segmentation model training method and apparatus, an electronic device and a storage medium are provided. The semantic segmentation model training method includes: acquiring a teacher semantic segmentation model that is pre-trained, the teacher semantic segmentation model including a first teacher network and a second teacher network, the first teacher network having structural characteristics of low depth and high width, and the second teacher network having structural characteristics of high depth and low width; processing a sample image based on the teacher semantic segmentation model to obtain a first segmentation map and a second segmentation map; and training a student semantic segmentation model that is lightweight according to the sample image, the first segmentation map and the second segmentation map, so as to obtain a target semantic segmentation model.