A remote sensing ship detection method based on a feature pyramid and distance constraint FCN

Abstract

The invention provides a remote sensing ship detection method based on a characteristic pyramid structure full convolutional neural network, and mainly solves the problems of mutual overlapping of horizontal contour bounding boxes and low coverage and detection rate in the prior art. The method comprises the following steps: 1, selecting and cutting a sample picture in an existing remote sensing data set, and dividing the sample picture into a training sample and a test sample; 2, training the full convolutional neural network by using an overall loss function composed of a category loss function with pixel distance frame minimum constraint and a shape loss function; 3, inputting a test sample into the trained full convolutional network, outputting the test sample as a characteristic matrix of each pixel point, and performing coordinate transformation to obtain coordinates of the contour bounding box; and 4, screening and combining the pixel contour bounding boxes to obtain a detectionresult image and a coordinate file. According to the method, the contour boundary frame with the angle can be generated for the remote sensing ship image, the detection rate is high, the boundary isaccurate, and the method can be used for extracting the ship target from the remote sensing image shot by the optical sensor.

Description

technical field [0001] The invention belongs to the technical field of image processing, and relates to a remote sensing ship detection method, which can be used to extract ship targets from remote sensing images captured by optical sensors. Background technique [0002] Ship detection in remote sensing images aims to replace tedious manual work, and use deep learning methods to obtain more accurate ship detection results compared to traditional parallel outline bounding boxes. The current ship extraction methods can be roughly divided into three categories. The first category is knowledge-based ship extraction methods, such as: threshold method, template matching method; the second category is based on traditional machine learning methods, such as: clustering, support vector machines, Bayesian classification, etc.; these two categories The ship extraction effect of the method is not ideal, and the overall accuracy is low. At present, more researchers are working on ship d...

AI Technical Summary

Problems solved by technology

Although this method takes advantage of the excellent characteristics of the convolutional neural network, it can extract discriminative features and extract the results of these feature ships, but when using this model for ship extraction, there are certain deficiencies: First, the method When it is used for ship extraction, it does not take into account the unbalanced distribution of ship scales in remote sensing images; secondly, when using the convolutional neural network for ship extraction, it cannot take into account the mutual coverage and overlap of parallel outline frames in dense scenes; the above two Points greatly affect the effect of ship detection, the ship outline bounding box extraction effect is poor, and the recall rate of ship extraction results is low

Due to the framework of the mainstream target detection model based on this method, it can only generate horizontal contour bounding boxes, and for dense and overlapping ship targets, the generated horizontal contour bounding boxes will interfere with each other when non-maximizing suppression is performed. Poor rate and accuracy

Images