Attention modules in convolutional neural networks for small object recognition

A problem of small object recognition is frequently encountered in biomedical and security systems. However, the detection of such objects is often complicated by presence of dense clouds or infrastructure objects. Results of using various attention mechanisms to improve accuracy in small objects segmentation with convolutional neural networks are presented in this paper. Modules of channel attention and spatial attention are considered. This approach allows one to effectively suppress less informative channels and image areas, while enhancing more informative channels and image areas. Meanwhile, weights of the attention modules are automatically adapted to the input data during training. An assessment of influence of the attention mechanisms in convolutional neural network architecture on the ability to suppress complex backgrounds (clouds and infrastructure objects) and segment small objects is performed. The results are presented in the form of tables with test metrics and figures with precision-recall curves, ROC curves and heatmaps showing an effectiveness of background suppression. The results obtained allow one to implement the described attention modules in the convolutional neural networks of any complexity and increase the recognition accuracy of objects of 10-40 pixels in size on a complex background.