Research On Inland River Ship Detection And Recognition Based On Convolutional Neural Network

China has a vast territory and abundant water transportation resources.With the development of China’s economy,the cargo throughput of inland waterways is increasing year by year,and more and more vessels are passing through inland waterways,so the risk of ship accidents in inland waterways is increasing year by year.Monitoring of inland waterways can not only timely detect illegal vessels and reduce the risk of accidents,but also rationally dispatch inland waterway transportation resources and improve the operation efficiency of vessels.In the monitoring system of inland waterway,the most basic and most important link is to identify and detect the ships to obtain the number and position of the ships in the current video.At present,most researches on ship identification and detection adopt artificial feature extraction method and traditional machine learning algorithm.The artificial feature extraction method cannot adapt to the complex and changeable environment,and is easily affected by the clutter in images or videos,resulting in the low accuracy and low efficiency of ship identification and detection.Based on the above,a classification model based on convolutional neural network for the identification task of inland waterway ships is proposed in this paper.At present,there is no ship open source dataset supporting ship classification model training,so it is impossible to fully train the ship classification model.Therefore,a simple dataset of ship classification and recognition is constructed,and a training method based on transfer learning is proposed.This method can take advantage of the powerful fitting and spatial representation capabilities of convolutional neural networks to fully learn various features of ships.The differences between models using transfer learning and models not using transfer learning is compared and analyzed in this paper.Finally,the effectiveness of this method is verified by comparing with traditional machine learning algorithms.Based on the above ship classification model,inland ship detection model based on convolutional neural network is further proposed in this paper.Several common detection algorithms are compared and analyzed firstly,then the defects of these algorithms in the task of ship detection is pointed out in this paper,that is,it is difficult to detect the occluded ships and small ships in the video.In the case of occlusion,a new loss function is used to optimize the model,and the weight attenuation strategy is used to replace the original zero resetting strategy in the process of post-processing.Then a comparison experiment is made between the new model and the original model in the dataset to verify the effectiveness of the method in the case of occlusion.For the small ship,the reason why it is difficult to detect the small ship is analyzed firstly.The feature pyramid method is used for information fusion to improve the information carried by small target ships in this paper.This method can fuse high-level feature maps with rich semantic information and low-level feature maps with rich spatial information to enhance the detection capability of small ships.Finally,through a comparison experiment with the original model,the effectiveness of the feature pyramid method in small ship detection is verified.