Deep Convolutional Network with Pixel-Aware Attention for Smoke Recognition

Deep convolutional networks have significantly improved the performance of smoke image recognition. However, the trained spatially-shared weights are applied to all pixels irrespective of the image content at the specific position, which may be suboptimal to address complicated smoke variants in shape, texture and color. Based on this background, we propose a deep convolutional network with pixel-aware attention for smoke recognition. A pixel-aware attention module is devised to modify the standard convolution in a pixel-specific fashion. The learned weights are dynamically conditioned on pixels in the smoke image, adaptively recalibrating the pixel features at the identical position along feature channels, and therefore enrich the feature representation space. Then, we build a simple and efficient deep convolutional network by introducing pixel-aware attention modules to recognize smoke images. Experimental results conducted on the publicly available smoke recognition database verify that the proposed smoke recognition network has achieved a very high detection rate that exceeds 98.3% on average, superior to state-of-the-art relevant competitors. Furthermore, our network only employs 0.3M learnable parameters and 90M FLOPs.