A Lightweight Cross-Layer Smoke-Aware Network

Abstract

Smoke is an obvious sign of pre-fire. However, due to its variable morphology, the existing schemes are difficult to extract precise smoke characteristics, which seriously affects the practical applications. Therefore, we propose a lightweight cross-layer smoke-aware network (CLSANet) of only 2.38 M. To enhance the information exchange and ensure accurate feature extraction, three cross-layer connection strategies with bias are applied to the CLSANet. First, a spatial perception module (SPM) is designed to transfer spatial information from the shallow layer to the high layer, so that the valuable texture details can be complemented in the deeper levels. Furthermore, we propose a texture federation module (TFM) in the final encoding phase based on fully connected attention (FCA) and spatial texture attention (STA). Both FCA and STA structures implement cross-layer connections to further repair the missing spatial information of smoke. Finally, a feature self-collaboration head (FSCHead) is devised. The localization and classification tasks are decoupled and explicitly deployed on different layers. As a result, CLSANet effectively removes redundancy and preserves meaningful smoke features in a concise way. It obtains the precision of 94.4% and 73.3% on USTC-RF and XJTU-RS databases, respectively. Extensive experiments are conducted and the results demonstrate that CLSANet has a competitive performance.