Research On Key Technologies For Ship Target Recognition Application

Under the conditions of the new era,the implementation of national strategies such as "building a strong ocean" and "the 21st century Maritime Silk Road of the Belt and Road" has become an inevitable choice for globalization.At the same time,disputes over marine resources and channel security have intensified,and maritime rights and interests are facing an intricate international situation.As the main body of various marine activities,how to identify the ship target in the ocean timely and accurately is the key to realize the national strategy of building a strong ocean.However,due to the particularity of the marine environment,it is difficult for the traditional target recognition methods based on deep learning to meet the application requirements under limited conditions such as computing resources,communication bandwidth,and energy supply.Therefore,focusing on the problem of automatic ship target recognition and the application of resource-constrained platform in the marine environment,this paper carries out research on key technologies related to lightweight ship identification based on deep learning.The main works are as follows:(1)Aiming at the problem of excessive model parameters in traditional ship-based target recognition methods based on deep learning,basing on the techniques of optimized convolution operation,model parameter compression and enhanced feature representation,the lightweight feature extraction structure is proposed to face resource-constrained platform applications.It has achieved high-quality extraction of target features in the case of drastically simplification of model parameter quantities and computational complexity.On the ImageNet-67 dataset,compared with the existing lightweight methods such as MobileNet,ShuffleNet,MobileNet v2,etc.,the comparison experiments were carried out in the aspects of recognition speed,accuracy,model parameter quantity,floating point calculation amount and memory occupation.The experimental results show that the overall performance of the proposed method is better than the existing ones.When model size is 7.8MB,it can achieve the technical indicators:a single frame time-consuming 16.3ms and classification accuracy of 93.5%.(2)Aiming at the problem of low utilization efficiency of traditional ship target recognition methods based on deep learning,comprehensive using of such techniques as ship target size clustering based on prior knowledge,high and low level feature fusion based on information fusion as well as the cross entropy loss function base on attention-based mechanism,an efficient method of using feature data has been proposed.Based on the VOC 2007 and the self-built ship target dataset,it compares with the SSD,DSSD and DSOD detection methods from the aspects of accuracy,memory consumption,real-time and model size.Experiments show that the comprehensive performance of the method mentioned in this paper is better than the existing method.Compared with the algorithm of SSD,the accuracy of 88.8% ship target recognition is realized under the condition that the storage space occupies only 4.02%.(3)For the energy saving problem in the application platform,taking into account the low speed of the ship target and the small change of the target between adjacent frames,the histogram-based video key frame extraction method is used to get the key frames in the video.And then only key frames are used for ship target recognition to improve the effective energy utilization of the application platform.(4)For resource-constrained platform applications,using Flask,Hadoop,Caffe and other tools,based on the embedded platform,the lightweight target detection prototype system was designed and implemented with real-time video ship target recognition and statistics.Finally,the instantiation test has been executed.