| Due to the problems of target missing detection and information loss of navigation radar,subjective unreliability of AIS report data and subjective negligence of manual lookout,which lead to frequent maritime safety accidents,exploring new navigation visual perception technology is an urgent problem to be solved in the development of intelligent ships.Computer vision technology has been widely used in all walks of life,but the application of vision technology in ship navigation visual perception is extremely difficult due to the wide and far space range of ship visual perception,poor hydroclimate and light conditions,and large differences in ship dimensions and shape.Relying on the national key research and development program,Safety Risk Identification,Prevention and Control Platform for Ships Underway,this paper conducts research on the problems of multi-scale target detection,visual enhancement under weak observation conditions,and difficulties in ship type identification under small sample conditions in the visual perception of ships underway,mainly including the following research contents.Firstly,for the problem of poor detection effect of current target detection methods on small targets of ships,a unified framework and algorithm for multiscale ship target detection is proposed by studying the visual feature enhancement method for very small targets and designing a multi-scale target deep learning network structure.The algorithm enhances the training data of small-scale targets by improving the data enhancement algorithm,which improves the training effect of small-scale targets.The attention module is added to the CSPDarknet network to construct the ACDarknet network,which promotes the multi-scale feature extraction network,increases the small target detection layer,and improves the detection rate of very small ship targets.Considering that targets in dense ship target areas are easily suppressed by the NMS,an adaptive NMS algorithm is designed to self-adaptively adjust the NMS thresholds of different areas to reduce the target missing detection rate in dense target areas.The experimental results indicate that the algorithm has significantly improved the detection effect of targets of different scales,with the algorithm target recall rate of 89.4%,and the detection accuracy of 93.6%.Secondly,in order to solve the problems of weak ship target features and easy for missing detection in rainy and foggy days,a self-adaptive ship target detection method is proposed.For the deficiencies of color offset and image darkness of traditional image enhancement algorithm,DefogCNN defogging network based on neural network is established.To improve the defogging effect of defogging network,a water surface fogging algorithm for training defogging network is designed.The feature differences between ship targets in sunny days and rainy and foggy days are analyzed,and the framework structure of selfadaptive ship target detection is designed.The experimental results show that the algorithm effectively improves the detection effect in foggy days,and the target recall and detection accuracy for the foggy day data set are 88.6%and 92.4%respectively.Thirdly,in view of the insufficient ship sample database and type recognition of ship target under small sample conditions,a ship type recognition network and a feature reconstruction module are designed to reconstruct the parameters of the ship target feature vectors obtained by the feature extraction network and enhance the small sample target learning capability of the classification network.The experimental results show that the detection accuracy of the recognition algorithm in the test set is significantly better than other algorithms,reaching up to 97.4%.The classification accuracy in the small sample data set is 89.2%,which is higher than other comparative alqorithms.Finally,a visual perception system for ships underway is developed.The system integrates visible light,infrared and laser ranging sensors and other major ship navigation equipment such as AIS,GPS,compass and radar information.By applying the multi-scale target detection method,ship target vision enhancement algorithm under rain and fog conditions,and ship type recognition technology under small sample conditions proposed in this paper,the accuracy and practical value of ship visual perception system in ship target detection and type recognition are effectively improved,which adds a pair of "eyes" to ships underway,makes up for the shortage of manual lookout and further ensures safe navigation of ships. |
