Ocean Remote Sensing Target Detection Based On Deep Learning And Internal Wave Simulation Imaging

With the rapid development of high-resolution optical remote sensing imaging technology,there is an urgent need for a rapid and reliable target detection technology to identify important targets in remote sensing images.At the same time,the sea target detection technology plays an important role in marine safety inspection and cargo transportation.However,there are still some challenges in marine target detection.Firstly,the targets in some areas of the ocean remote sensing image are dense and non oriented,which easily leads to false detection and missing detection of the model;Secondly,advanced remote sensing devices usually produce high-quality images of several thousand pixels,which contain a lot of irrelevant information,making the detection of specific targets complicated;Finally,the training of the model needs a lot of data support.In the military application,with the development of stealth technology,it is difficult to obtain the data directly.This paper aims to build a target detection model for marine area based on deep learning.Firstly,based on the open dota data set,the initial ocean data set is obtained by data segmentation.At the same time,considering the difficulty of submarine target data acquisition,according to Kelvin wake hydrodynamic model,this paper calculates the geometric development of simple submarine wake under different working conditions,and fuses the submarine wake with the accident ocean data set by image fusion,and completes the construction of ocean data set by constructing tags.In order to reduce the influence of irrelevant information in the ocean,fast MBD saliency fusion algorithm is used to obtain the saliency image of the data set,and the original data set is used to share tags for training.At the same time,in order to increase the robustness of the model,this paper uses a variety of data enhancement methods,and introduces the k-means algorithm,and finally completes the data preprocessing before the model training.For the construction of ocean remote sensing target detection model,based on the model of yoov4,the main network is lightweight to reduce the difficulty of model fitting,and the lower layer is selected as the output layer of the model to enhance the ability of small target detection;Then the angle variable is introduced to solve the problem of arbitrary direction of the target in remote sensing image;At the same time,it enhances the fitting ability of the model to the angle.This paper improves the NMS algorithm to calculate the actual intersection and union ratio between the two inclined frames;Finally,the SPP structure of the model is improved and the receptive field of the model is enlarged.After fully training the model,this paper evaluates the performance of the model.Firstly,the ablation experiments are carried out to study the effects of RNMS,K-means algorithm,image saliency fusion and data enhancement on the model detection;Then,this paper studies the data set cutting method,and analyzes the influence of the size and span of the cut image on the detection effect;Secondly,the submarine wake data set is constructed to verify the detection ability of the model;Finally,in the same environment,the model is compared with other advanced models to prove the detection ability of the model.