MFF-Net: A multitask feature fusion network in dual-frequency domains for detecting smoke from one single picture.

The industrial smoke detection is extreme significant for sustainable development and human health, but up to date, its related technologies have not met the real-time and accurate requirements, particularly in the memory-constrained conditions. Towards detecting smoke from one single picture more accurate and efficient, in this paper we present a multitask feature fusion network, dubbed MFF-Net. To specify, according to the characteristics of imbalance found in the frequency distribution of numerous smoke pictures, our MFF-Net is set up with a dual-channel framework, which is composed of the high-frequency and low-frequency channels that process different frequency components respectively. The high-frequency channel is built based on a stream of cross-level fusion inverted residual (CFIR) blocks, each of which encompasses the expansion layer, filtering layer, combination layer, and fusion layer, while the low-frequency channel is formed by embedding the attention mechanism into CFIR blocks. We concatenate the dual channels mentioned above to establish the MFF-Net, in order to systematically fuse high- and low-frequency features of a single smoke picture. As compared with the existing relevant detection models, the depthwise and pointwise convolutions are utilized to make networks more efficient with respect to size and speed, while the dense skip connection is used behind to fuse cross-level features (i.e., semantic and visual features) for keeping high accuracy. Experimental results show that our proposed MFF-Net can trade off multiple tasks in accuracy, memory and computational efficiency. • Smoke pictures have characteristic of imbalance frequency distribution. • A fusion network is proposed in dual-frequency domains for detecting factory smoke. • Novelty 4-layer blocks encompass expansion, filtering, combination & fusion layer. • Our model achieves accuracy of 99.94% and only needs 0.59 M-params & 1.16 M-flops. • Our model trades off multitasks, i.e., accuracy, memory & computational efficiency. [ABSTRACT FROM AUTHOR]