Research On Few-shot Ship Object Detection Based On Generative Adversarial Networks And Transfer Learning

The area of the territorial sea of China is vast,accounting for one third of territorial area and bordering some other countries.In recent years,the situation of maritime security has been increasingly complex.There is no facts in denying that some countries covet ocean resources of our country.Meanwhile,The news that foreign ships intrude our territorial waters has frequently reported by medias.Considering the reason for bitter military disputes,ensuring the safety of the territorial sea is crucial.Therefore,it is a very important task that relevant institutions strengthen surveillances and real-time detection of invading ships in our waters.Nevertheless,the methods of traditional ship target detection result in the limitation of ship target detection performance because of just using the shallow information.Therefor,It is necessary to study the reliable method of robust ship target detection.As the development of the approach of deep learning improves the level of the fine observation.The possibility of deeply developing digging into target information and effectively using the ship target detection method of target fine information has been exploited.However,as the military research background and ship target image is not easy to acquire,studying the ship target detection method with small samples is urgent for improving the ship target detection ability and accuracy.This paper focuses on the above problems.the main research contents and innovation points are as follows.(1)Aiming at the problem of insufficient and fuzzy representation for the image feature obtained by self-attention-generated countermeasures network(SAGAN),a small sample data enhancement method based on the improvement for the self-attention-generated countermeasures network(SAGAN)is proposed in this work.Firstly,by adding additional conditional features into the generator and discriminator,and the dimension information of the data is correlated with the semantic features,the robust indication for the category information is achieved.Then,the Adam optimizer is introduced to optimize the parameters in the initial phase for the parameters training,and the appropriate learning rate is set based on the prior knowledge.Finally,the training situation and test output results are analyzed on the trained detection model.Comparing with the existing SAGAN method,the improved SAGAN method has advantages in the clarity and diversity for the ship samples.The effectiveness of which is verified by the comparison of experimental results.(2)Aiming at the over-fitting and difficulty for the adjusting of parameters accurately in traditional deep learning network model with small samples,a method based on transfer learning for ship target detection is presented.Firstly,an adaptive fine-tuning network module using competitive strategy is introduced.The convolutional neural network VGG-16 and Google Net are improved to enhance the performance of pre-training model.And then,the pre-trained VGG-16 and Goog Le Net network model feature parameters are migrated to Faster R-CNN and YOLO networks.Finally,the improved ship target samples data are utilized to fine-tune,train and test the network to realize the robust detection of ship target with small samples.Compared with the existing detection methods,the proposed method can accurately estimate the optimal training depth and characteristic parameters of the pre-training network and solve the problem for low performance of over-fitting and detection under small samples.The effectiveness of the proposed method is verified through experiments.