Teeth attention mechanism image classification based on lightweight network

Abstract

In order to solve the problem that embedding attention mechanism module into lightweight convolutional neural network will increase the number of parameters and FLOPs with less accuracy improvement, a Teeth Attention Module (TAM) based on lightweight network was proposed. The feature map is enhanced from two aspects of channel attention and spatial attention. In the channel attention module, local channel interaction is carried out from height and width without dimensionality reduction through one-dimensional convolution, and it is divided into MP-ECAM and AP-ECAM according to pooling mode. In the spatial attention module, the average value and maximum value of each pixel point in the feature map are calculated according to the dimensions to construct ESAM. Finally, TAM is constructed in a channel-space-channel manner. TAM was embedded into the lightweight networks MobileNetV2, ShuffleNetV2 and EfficientNetV1 respectively, and the experiment was carried out on CIFAR-10 image classification datasets. Compared with Squeeze and Excitation Module (SE), Convolutional Block Attention Module (CBAM) and Efficient Channel Attention Module (ECA), this module can achieve higher accuracy with less increase of the number of parameters and FLOPs.