A Gated Recurrent Network With Dual Classification Assistance for Smoke Semantic Segmentation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Smoke segmentation is challenging due to its semi-transparent nature, creating complex mixtures with backgrounds.
• Sparse or small smoke is often inconspicuous with ambiguous boundaries, making detection difficult.

Purpose of the Study:

• To develop an advanced deep learning model for accurate smoke semantic segmentation from single images.
• To address the challenges posed by inconspicuous smoke and smoke-like objects in image analysis.

Main Methods:

• Proposed a Classification-assisted Gated Recurrent Network (CGRNet) incorporating dual classification assistance for smoke discrimination.
• Introduced an Attention Convolutional GRU (Att-ConvGRU) for capturing long-range feature dependencies.
• Designed Multi-scale Context Contrasted Local Feature (MCCL) and Dense Pyramid Pooling Module (DPPM) to enhance feature representation for small smoke detection.

Main Results:

• The CGRNet achieved significant performance improvements at the image level through dual classification assistance.
• The Att-ConvGRU, MCCL, and DPPM modules enhanced the network's ability to perceive small and inconspicuous smoke.
• Demonstrated superior performance compared to state-of-the-art algorithms on smoke datasets.

Conclusions:

• The proposed CGRNet effectively tackles the complexities of smoke segmentation in single images.
• The method shows robust performance, even on challenging images containing inconspicuous smoke and smoke-like objects.