Instance segmentation convolutional neural network based on multi-scale attention mechanism.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Instance segmentation is crucial for scene understanding in robotics, autonomous driving, and medical imaging.
• Existing methods struggle with low-resolution objects and complex backgrounds, limiting detection efficiency and speed.
• There is a need for advanced instance segmentation techniques to overcome these limitations.

Purpose of the Study:

• To propose a novel instance segmentation method, Hybrid Kernel Mask R-CNN, addressing efficiency and speed challenges.
• To enhance the detection of low-resolution objects and improve performance on complex scenes.
• To achieve state-of-the-art results in instance segmentation.

Main Methods:

• Developed a Hybrid Kernel Mask R-CNN incorporating hybrid convolution kernels for rich information extraction.
• Implemented a multi-scale attention mechanism to assign weights to convolution kernels, retaining crucial information.
• Focused the network on low-resolution objects within images.

Main Results:

• The proposed Hybrid Kernel Mask R-CNN achieved superior accuracy compared to anchor-based methods.
• The method demonstrated state-of-the-art performance on Balloon, xBD, and COCO datasets.
• Visualization confirmed effective extraction of low-resolution objects.

Conclusions:

• Hybrid Kernel Mask R-CNN offers a significant advancement in instance segmentation.
• The method effectively addresses challenges with low-resolution objects and complex backgrounds.
• The approach shows broad applicability and superior performance across diverse datasets.