Deep Learning-Based Benthonic Organism Detection: Fuzzy Channel--Spatial Attention
提出一种模糊通道-空间注意力机制，用于增强底栖生物特征并抑制水下背景噪声，在深度学习框架中集成该模块，实验表明其性能优于现有水下检测方法。
In this article, to augment benthonic organism features and suppress underwater background noises, simultaneously, a fuzzy channel–spatial attention-based benthonic organism detection (FCSA-BOD) scheme is proposed. Main contributions are as follows: 1) with the aid of spatial global average and maximum pooling, fuzzy channel attention (FCA) is originated to adaptively recalibrate channel responses by fusing discriminative and textural channel attention maps, which are derived from two independent single-hidden-layer feedforward networks, such that benthonic organism and background feature maps can be strengthened and suppressed, respectively; 2) by exploiting channel global average and maximum pooling, fuzzy spatial attention (FSA) is created to highlight spatial regions associated with benthonic organisms by fusing multiple spatial attention maps possessing completely different receptive fields, such that different-scale benthonic organism features on the same feature map can be significantly augmented, simultaneously; and 3) the FCSA-BOD scheme is eventually established in a modular manner by integrating FCA and FSA modules within a deep learning framework. Comprehensive experiments demonstrate that the proposed FCSA-BOD scheme outperforms state-of-the-art underwater detection approaches.