A small target detection method based on global background perception and feature enhancement

Abstract

The application discloses a kind of small target detection methods based on global background perception and feature enhancement, belong to image target detection technical field.The method in this paper: first, the deep feature map of input image is extracted by pre-training backbone network, then global background feature is obtained after global average pooling;Second, the small target candidate region feature is extracted using region proposal network, and is up-sampled by deconvolution, and then the small target feature is further enhanced using enhancement convolution block to obtain more rich target feature;Then the small target enhanced feature and global background feature are spliced to form the final output feature, and then the target positioning and classification task is carried out through fully connected layer;Finally, the difference between small target and background feature is optimized by designing contrast loss function.The method can improve the performance of small target detection by global background perception and feature enhancement, and is suitable for small target detection task in complex background.

Images