Method for dim and small object detection based on discriminant feature of video satellite data

A method for detecting a dim and small object based on a discriminant feature of video satellite data, comprising:

• S1:
• cutting an image obtained from video satellite data into image blocks with a same size and inputting the image blocks into a VGG16 backbone network to obtain an image feature of the video satellite data;
• S2:
• performing data annotation on the cut image blocks, taking an annotation region obtained by the data annotation as a range for extracting a discriminant feature, adopting a symmetric semantic segmentation model to obtain a segmented image with pixel-by-pixel semantic annotation and extracting an image feature of an object that is significantly different from a background to obtain the discriminant feature of the object;
• S3:
• fusing the image feature of the video satellite data and the extracted discriminant feature of the object by utilizing a top-down adjustment mechanism;
• S4:
• by utilizing an attention mechanism, performing background enhancement on a background of a segmented image obtained after segmentation by the symmetric semantic segmentation model in step S2 to further increase contrast between the object and the background;
• and
• S5:
• performing multi-scale fusion on fusion features with different scales in the image obtained by the background enhancement in S4, to extract the dim and small object in the video satellite data by introducing a multi-scale semantic analysis strategy and adopting a pyramid model;
• wherein, S5 comprises:
• S51:
• simulating a perception process from coarse to fine in a human visual system by the multi-scale semantic analysis strategy;
• wherein for an object detection of the video satellite data, high-scale data are used to detect small objects, and low-scale data are used to detect large objects;
• S52:
• constructing a multi-level feature pyramid structure and a feature pyramid, wherein high-level features are conducive to classification, and low-level features are conducive to regression of a position of the object, and the feature pyramid is stacked by adopting the multi-level feature pyramid structure to provide a multi-level feature pyramid connection structure;
• and
• S53:
• encoding and decoding the discriminant feature by using a U-shaped network, weighting and splicing feature channels with a same scale among different layers in the feature pyramid connection structure, and performing multi-scale prediction based on the fusion features obtained in S3 to detect the dim and small object in the video satellite data.