Research On Detection Of Water Surface Object Image Based On SSD Algorithm

As China has proposed a strategy of strengthening the country by ocean,transportation will be busy progressively.In the meantime,under the impact of a new background of artificial intelligence,the intelligentization of water transportation has become the trend from now.As an intelligent marine equipment,unmanned boats will accelerate the development of the marine economy rapidly.The current unmanned boat technology mainly relies on radar,sensors and traditional object detection technology.With the development of computer hardware technology,the deep detection-based object detection technology can mine the deep feature information of the image,which greatly compensates for the traditional object detection and the lack of radar technology.For another purpose it also can save a lot of cost.This paper will use the deep learning method to study the surface object detection.This paper comprehensively analyzes the detection speed and detection accuracy of today's classical algorithms which based on deep learning.The SSD algorithm is used to design the network structure of water surface object detection.The training,verification and testing are carried out through self-collecting,labeling and standardized water surface target data sets.The test results were compared and analyzed.Through comparative analysis,it was found that the surface object detection based on SSD algorithm is not ideal for detecting small target data and often interrupted when training data.In order to solve the inaccurate positioning of the water surface object detection of SSD algorithm and the difficulty of extracting features of small target,this paper uses K-medoids algorithm to optimize the aspect ratio of the default candidate frame.So that the aspect ratio of the default candidate frame is more Fit the surface target dataset to reduce the number of positional offset regressions during model training.In order to better extract the characteristics of small targets and increase the number of small target data sets,this paper proposes a new data enhancement method-multi-object combined images,increasing the number of training sets for small targets.In order to solve the problem that the SSD algorithm is interrupted during training,the Elu function is used to replace the original Relu activation function of the SSD algorithm to improve the training efficiency of the model.In order to improve the training speed of the surface object detection model,a pre-training model is added and the parameters are fine-tuned appropriately.Through experimental comparison,the quantitative and qualitative analysis of the water surface object detection based on SSD algorithm is carried out before and after optimization.The detection result of the optimized model improves the detection effect of the SSD algorithm on small targets to some extent and is not interrupted during model training.The optimized model has an increase of 2.2 percentage points in average detection accuracy compared with that before optimization.Therefore,the optimized SSD algorithm-based water surface object detection model is superior to the traditional SSD algorithm-based water surface object detection model.