Image Segmentation Using Encoder-Decoder with Deformable Convolutions.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Image segmentation is crucial for image analysis but faces challenges with real-world image complexities like noise and obstruction.
• Existing approaches often lack a universally applicable solution.

Purpose of the Study:

• To propose a novel semantic segmentation architecture using a modified Xception convolutional neural network.
• To enhance the model's performance through experimentation and the integration of a deformable convolution module.

Main Methods:

• Utilized the Xception model as a base for a convolutional neural network architecture.
• Conducted experiments varying network resolution, depth, and employing data augmentation techniques.
• Integrated a deformable convolution module to improve feature representation.

Main Results:

• Achieved a 76.8 mean IoU on the Pascal VOC 2012 dataset.
• Obtained a 58.1 mean IoU on the Cityscapes dataset.
• Outperformed SegNet and U-Net in performance while requiring fewer parameters and less inference time.

Conclusions:

• The proposed Xception-based architecture with deformable convolutions offers a superior solution for semantic image segmentation.
• This approach provides a more efficient and effective method for image analysis compared to existing networks.