Anti-Attention Mechanism: A Module for Channel Correction in Ship Detection
Qiu Xinjie, Fenglei Han, Wangyuan Zhao · 2023
The difference of ship image style will lead to poor effect of ship detection. Aiming at the characteristics of the singleness of the dataset's image style in ship detection, we propose an anti-attention module for correcting detection results at the neural network level, in order to make the detection network pay less or less attention to some feature layers. Specifically, firstly, an original feature layer is input into two additional trained convolutional neural networks to get two new weights; secondly, the two output weights are filtered through certain parameters to screen out the feature layers that are not conducive to detection, and multiplied by a certain attenuation coefficient; finally, the output results are weighted so as not to affect the overall result. After the corresponding experiments, only the correction effect of a single anti-attention mechanism improves the results by 1.3% and 1.7% in F1-score and [email protected] before correction. Therefore, the anti-attention mechanism we propose has a certain corrective effect in the detection stage under certain circumstances.