Research On Target Separation And Intelligent Recognition Of Maritime Images Under Fog

Fog weather will have a significant impact on the safety of ships navigation.The installation of visual sensors on the ship will help to know and judge the maritime situation in advance and improve the safety of autonomous navigation of smart ships.However,in foggy conditions,images acquired by sensors on the sea often have problems such as low contrast,color distortion and detail loss.In addition,due to the interference of sea spray and reflection,the feature information such as texture and shape of sea image is highly interfered.Therefore,in order to improve the ability of automatic detection and identification of intelligent navigation system of ships in fog environment and reduce the rate of missed detection and false alarm.This paper focuses on the problem of object separation and intelligent identification of maritime images in fog.Through image enhancement,sea antenna segmentation,salient region detection and separation,and target intelligent identification,a scientific and effective method for target separation and identification in foggy marine image using vision sensor is systematically established.Firstly,the image enhancement algorithm based on Retinex model is improved according to the imaging characteristics of images under fog.By analyzing the image dehazing algorithm of traditional Retinex model,a new Retinex image enhancement algorithm based on Gaussian pyramid transform based on improved bilateral filtering was proposed,aiming at the problems of the single-scale Retinex model after image enhancement,such as the difficulty in highlighting local details and poor image contrast.Resolve image detail blur and improve image contrast.In order to solve the problem that the multi-scale Retinex(MSRCR)model with color restoration has the overall color distortion and poor visual effect after image enhancement.An MSRCR image enhancement algorithm based on global brightness adaptive equalization is proposed to provide more abundant color,texture and edge image features for the subsequent image processing.Secondly,a fast detection algorithm of sea-sky line based on gradient integral and polynomial iteration is proposed to achieve fast segmentation of sea surface and sky.Gaussian low pass filter is used to enhance the gradient edge features of the image,and the gradient integral values in the given region are sliding statistics to determine the potential region of the Marine antenna in the sea image.The maximum gradient points in the potential area were found as the candidate points of the sea antenna in each column,and polynomial iterative fitting was carried out for all the candidate points.After eliminating the false detection points,the final sea antenna was determined by fitting,so as to obtain images containing sky background and sea surface images containing target objects.Thirdly,an object separation algorithm of visual attention mechanism is proposed which fuses multiple visual features of images.Synthetic saliency map is obtained to segment salient target region of maritime image.Eight visual feature sub-images were selected,including frequency feature,direction feature of improved Gabor operator,gradient texture feature,brightness feature and color antagonism feature extracted from color space transformation.The salient sub-graphs of different visual features are obtained by spectral residual algorithm.The significance density function was used to calculate the weight coefficients of each salient sub-graph.A comprehensive salient map with multiple visual features was constructed.After the adaptive threshold segmentation of the comprehensive salient map,the salient regions in the comprehensive salient map were separated and obtained by the self-growing strategy of salient regions.Finally,An intelligent identification and classification algorithm of Marine image targets based on convolutional neural network framework is studied to complete the classification and intelligent identification of Marine targets.Combining the advantages of VGGNetwork and Res Net network,a new CNN framework was constructed to train and learn the training data set containing 96980 images in 9 categories.And 50 images of each category in the data set are randomly selected for data enhancement as the test data set of intelligent identification experiment.The experimental results show that the algorithm has higher training speed and lower training error.The average time of a single training iteration was 0.9s.The identification accuracy of the test data set was 95.14%.In addition,a comparative experiment of intelligent identification of significant area image sets based on the visual attention mechanism of fog maritime image target separation is completed.The experimental results show that the proposed algorithm has advantages in recognition speed,and the identification time of a single image is faster than VGG-16 algorithm by 86.6%,and faster than Res Net50 algorithm by 24.35%.At the same time,the maximum image recognition rate is better than Res Net50 algorithm by 25%,and slightly lower than VGG-16 algorithm by 2.78%.