SE-Attention是一种增强神经网络对输入特征图关注的注意力机制,使网络能够更好的捕获重要通道的信息[26],其网络结构如图2所示.首先对每个通道执行全局平均池化(global average pooling)操作,将每个通道的特征图压缩(squeeze)到特定的标量值,以表示该通道在整个特征图上的平均重要性.其次在激励(excitation)阶段,SE-Attention引入1个小型的前馈神经网络,将多个通道的标量值整合为单个向量后,经过全连接层(FC)的处理之后,再利用sigmoid函数将其变换为0 ~ 1之间的值,并对特征图的通道维度进行重新加权,以强调重要的通道,并抑制不重要的通道.其重要思想在于允许不同通道之间的交互,通过在Squeeze阶段对每个通道进行全局平均池化,网络可以捕获通道之间的关系,从而能够更好地学习到哪些通道在特定任务中更为重要,从而增强模型对特征图的关注,有助于提高模型的表征能力和泛化能力.自适应的学习通道的权重,计算成本较小的优势也使得SE-Attention更便于模型嵌入,从而提高模型性能.
BOCKC H, POOLEG H, PARKERP E, et al. Plant disease severity estimated visually, by digital photography and image analysis, and by hyperspectral imaging[J]. Critical Reviews in Plant Sciences, 2010, 29(2): 59-107.
[6]
ARNAL BARBEDOJ G. Digital image processing techniques for detecting, quantifying and classifying plant diseases[J]. SpringerPlus, 2013, 2(1): 660.
WANGY, DAIY, XUEJ, et al. Research of segmentation method on color image of Lingwu long jujubes based on the maximum entropy[J]. EURASIP Journal on Image and Video Processing, 2017, 2017: 1-9.
BAIX, LIX, FUZ, et al. A fuzzy clustering segmentation method based on neighborhood grayscale information for defining cucumber leaf spot disease images[J]. Computers and Electronics in Agriculture, 2017, 136: 157-165.
REDMONJ, DIVVALAS, GIRSHICKR, et al. You only look once: Unified, real-time object detection[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2016: 779-788.
GUOM H, XUT X, LIUJ J, et al. Attention mechanisms in computer vision: a survey[J]. Computational Visual Media, 2022, 8(3): 331-368.
[26]
DENGJ, MAY, LID, et al. Classification of breast density categories based on SE-Attention neural networks[J]. Computer Methods and Programs in Biomedicine, 2020, 193: 105489.
[27]
CHENJ, MAIH S, LUOL, et al. Effective feature fusion network in BIFPN for small object detection[C]//2021 IEEE International Conference on Image Processing (ICIP). IEEE, 2021: 699-703.