Accurate detection method for dense ship targets based on high-resolution remote sensing images

Abstract

The invention relates to an accurate detection method for dense ship targets based on high-resolution remote sensing images. First, the features of a remote sensing image are extracted through a convolution neural network, and then, the features are fused through up-sampling and the convolution features extracted by the convolution neural network. Target prediction is carried out independently ateach point on the feature map obtained by feature fusion. Specifically, the score of probability that each point on the feature map belongs to a target, the distance from the point to each of the foursides of a target frame where the point is located and the angle of the target frame where the point is located are predicted at the same time. When the score of probability that a point on the feature map belongs to a target is greater than a set threshold, a detected target frame can be calculated according to the distance from the point to each of the four sides of the target frame where the point is located and the angle of the target frame where the point is located. As target prediction is carried out independently and intensively at each point on the feature map, a final target detection result can be obtained through non-maximum suppression on the predicted target frame.

Description

technical field [0001] The invention belongs to the technical field of computer vision, and relates to a method for accurately detecting ship targets in high-resolution remote sensing images, specifically a method for accurately detecting ship targets in high-resolution remote sensing images based on a fully convolutional neural network (FCN) framework method. Background technique [0002] Target detection is an important and challenging task in computer vision. In recent years, target detection based on conventional natural images has made significant progress. Cutting-edge target detection algorithms (such as Faster R-CNN, Yolo, SSD, MaskR-CNN etc.) are all experimented on regular natural image datasets. [0003] Different from conventional natural images, ship targets in remote sensing images have their particularities, such as scale diversity, perspective diversity, small target density problems, multi-directional problems, and high background complexity. If they are th...

AI Technical Summary

Problems solved by technology

[0003] Different from conventional natural images, ship targets in remote sensing images have their particularities, such as scale diversity, perspective diversity, small target density problems, multi-directional problems, and high background complexity. If they are the same as conventional data sets, The ship target in the remote sensing image is detected in the way of horizontal frame labeling. When the ship target is relatively dense, the IOU (Intersection of Union) of the truth frame of the adjacent ship target will be relatively large, resulting in Object detection frameworks for natural images do not perform well on high-resolution remote sensing images

Images