Research On Water Target Recognition Based On Infrared And Visible Images

In the case of more and more shipping collections,the frequent occurrence of water accidents has caused serious harm to people’s lives and the ecological environment.The all-day intelligent recognition of water targets has become a common concern in countries around the world.With the continuous improvement of the performance of various sensing devices,multi-source images have a huge advantage in the acquisition of target information.Therefore,it is urgent to use different detection methods to realize the intelligent recognition of water targets and to maximize the safety of navigation.This paper mainly studies the methods of ship recognition based on infrared and visible images.The main contents include the following.Aiming at the problem of low accuracy of current ship recognition algorithms and the recognition accuracy of multi-form ships of different sizes in the same field of view,an accurate recognition method of polymorphic ships based on single-step cascade neural network is studied and implemented.The network consists of three modules: feature extraction,scale transformation and classification regression.First,cluster analysis of selfbuilt data sets to generate appropriate anchors as a priori frame for target detection;and introduce GIo U evaluation indicators during model training to improve the contribution of difficult samples to loss optimization;and finally design four-scale feature pyramid,Through the scale transformation module to complete the fusion of high and low-order features,and use the classification and regression module to complete multi-scale prediction;finally achieve the integrated and accurate identification of small ships and large ships under visible light conditions.Through the test of public data sets and self-built data sets,it is proved that this method is superior to other comparison methods in detecting recognition accuracy and speed.Aiming at the problems of insufficient paired infrared training data sets and difficult labeling in the process of infrared and visible light fusion recognition,a high-fidelity conversion method of infrared and visible light images based on a two-way generative adversarial network was studied and implemented.In this method,two part of generators and discriminators are designed to learn against each other,and iterative update of the network is realized through cross loss.At the same time,a depth residual block is introduced into the generator to complete the feature extraction and conversion of different levels of infrared and visible light images,and finally the mutual conversion and automatic annotation of infrared and visible images are completed,which provides high-quality training sets for subsequent fusion recognition.In order to solve the problem that the lack of guiding fusion strategy in the process of infrared and visible light fusion recognition leads to low fusion recognition accuracy,research and implement an accurate infrared and visible light ship recognition method based on symmetric fusion network.This method uses the dual-stream feature extraction module to complete the feature extraction of infrared and visible light images,and completes the feature fusion through concat.In the process of classification and regression,an adaptive weight migration module is designed to adjust the feature ratio of infrared and visible light to obtain the final detection and recognition results to complete all-day target recognition tasks.Testing with multiple sets of actual data prove that the recognition effect of this method is better than other methods.