Instance Segmentation Based on Improved Self-Adaptive Normalization.

Area of Science:

• Computer Vision
• Deep Learning
• Image Segmentation

Background:

• Batch Normalization (BN) is standard in instance segmentation but suffers from accuracy drops with small batches and GPU memory issues with large batches.
• Self-Adaptive Normalization (SN) offers improvements but faces challenges in weight learning and assignment due to its parameter averaging mechanism.

Purpose of the Study:

• To address the limitations of existing normalization methods in instance segmentation.
• To propose a novel adaptive weight loss layer and weight learning method for improved feature expression and batch size independence.

Main Methods:

• Replaced single Batch Normalization (BN) with an adaptive weight loss layer within Self-Adaptive Normalization (SN) models.
• Developed a new weight learning method to enhance input feature expression for subsequent layers.
• Implemented and validated the proposed method using a PyTorch deep learning framework on MS-COCO and Autonomous Driving Cityscapes datasets.

Main Results:

• The proposed method demonstrates effectiveness independent of batch size, ensuring stable accuracy for diverse target segmentation tasks.
• Achieved significant reduction in overall loss values.
• Improved the convergence speed of the deep learning network.

Conclusions:

• The novel adaptive weight loss layer effectively overcomes the drawbacks of traditional BN and SN methods in instance segmentation.
• The proposed approach offers a robust solution for instance segmentation tasks, particularly those constrained by batch size limitations.
• This method enhances network performance by improving feature expression, reducing loss, and accelerating convergence.