An ENet Semantic Segmentation Method Combined with Attention Mechanism.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Image semantic segmentation is crucial for applications like autonomous driving and medical imaging.
• Existing methods struggle with feature map fusion and lose detail due to downsampling.

Purpose of the Study:

• To develop an improved semantic segmentation algorithm addressing feature fusion and accuracy loss.
• To enhance the integration of channel and location-specific features within feature maps.

Main Methods:

• Utilized dilated convolution with a smaller downsampling factor to preserve image resolution and details.
• Integrated an attention mechanism module to assign weights to feature map regions, reducing accuracy loss.
• Designed a feature fusion module to merge features from different receptive fields effectively.

Main Results:

• The proposed method compensates for downsampling-induced accuracy loss while maintaining receptive fields and improving resolution.
• Experimental validation on Camvid, Cityscapes, and PASCAL VOC2012 datasets demonstrated significant performance improvements.
• Achieved higher Mean Intersection over Union (MIoU) and Mean Pixel Accuracy (MPA) compared to traditional methods.

Conclusions:

• The attention-based semantic segmentation algorithm effectively integrates multi-scale features and enhances segmentation performance.
• The method offers a robust solution for computer vision applications requiring high-accuracy image segmentation.
• The proposed feature fusion module significantly improves the model's ability to learn and integrate diverse features.