Visual information maritime target identification multi-source mixed data set construction method

Abstract

The invention belongs to the technical field of computer vision, and particularly relates to a visual information maritime target identification multi-source mixed data set construction method. According to the method, types of maritime targets are divided in detail, so that standard classification conditions are provided for subsequent sample collection and labeling; sample data is acquired by adopting multiple ways, and sample data sets are respectively constructed by utilizing modes of maritime target sample acquisition, network search and data sample augmentation; and a sample distribution and processing strategy is designed, and a maritime target identification evaluation strategy is formulated, so that composition proportions of various data in samples are more reasonable, and the accuracy of subsequent neural network training is ensured. According to the invention, more reasonable training samples can be provided for a maritime target identification network training task based on deep learning, so that a trained network has relatively high robustness and identification accuracy.

Description

technical field [0001] The invention belongs to the technical field of computer vision, and in particular relates to a method for constructing a multi-source mixed data set for visual information maritime target recognition. Background technique [0002] During the sailing process of the ship, the visual sensor can be used to obtain the surrounding image information, and then the deep learning network can be used to train such samples. The trained network can identify the target ship in the real-time collected sea target images online, and give information such as the type of the target ship , to provide the ship operator with the information of other ships in the image quickly, and improve the safety of ship navigation. However, the deep learning network needs to be based on a large number of training samples, so it is of great significance to construct a multi-source mixed data set for visual information maritime target recognition. [0003] At present, the typical constr...

AI Technical Summary

Problems solved by technology

[0004] The above literatures have the following problems in the process of information fusion: First, most of the data sets based on visual images taken by drones or real ships cannot cover all navigation environments, including day, night, cloudy, sunny and other environments, so The deep learning network trained by the data set constructed in this way is often only more accurate in target recognition in this environment, but the recognition effect in other cases is not good.

The second is that there are various types of sample images collected through the network, such as real scene images, simulation images, virtual scene images, etc., but the proportions of various types of data are not explained. The different proportions of various ship samples will affect the neural network training. and the effect of using

Images