Research On The Balanced And Complete Construction Technology Of Image Target Dataset

Deep learning has three core elements: big data,deep learning algorithm design and high-performance computing platform.In the field of computer vision target recognition,big data is embodied as a target dataset composed of big data of target image samples to be recognized.The target dataset is the premise and foundation of the target recognition algorithm based on deep learning,and its completeness,balance and scale will directly affect the performance of the algorithm.However,in the process of constructing the target dataset,affected by many factors such as the imaging conditions of the target to be identified,acquisition conditions,cooperative/non-cooperative targets,cost,etc.,the target dataset exhibits imbalance and incomplete features,which cannot meet the practical application needs.It is usually expressed as: "A well-trained deep network model can achieve a good target recognition effect in a laboratory environment,but the target recognition effect is extremely poor in an actual application environment".At present,most of the research on target datasets focuses on the construction of public datasets,the problem of unbalanced datasets,and the optimization of the balance between target classes based on deep learning networks.Little attention is paid to the balance and completeness of the construction of specific target datasets in actual scenarios.Therefore,under the actual application requirements of ship target recognition,this article carry out research on key technologies related to the balanced and complete construction of ship image target datasets,and improve the accuracy and robustness of the ship target recognition algorithm in practical engineering applications.The main work content as follows:(1)The concept of balance and completeness of target dataset is proposed.We combine with actual ship target recognition application scenarios,and analyze the completeness and balance connotation of ship target dataset in detail.Finally,we propose a ship target dataset evaluation index system,which can provide support for the balanced and complete evaluation of ship target dataset.(2)A method for evaluating the balance and completeness of target datasets is designed.The method is designed based on the positive correlation between the completeness of the target dataset and the accuracy of target recognition,that is,the higher the completeness of the target dataset,the better the accuracy of target recognition.First,we perform statistical analysis,processing,and calculation on the target dataset to obtain the evaluation index parameters;second,we calculate the correlation and weight between the evaluation index parameters of the target dataset and the target recognition accuracy;third,the optimal/worst balanced and complete solution for the target dataset is generated;fourth,we perform a balanced and complete evaluation of the current target dataset.The self-built ship target dataset is used to verify the balance and completeness evaluation method of the target dataset.The experimental results explain that the method can not only quantitatively estimate the balanced and complete performance of the ship target dataset,but also give a qualitative direction to improve the balanced and complete performance of the ship target dataset.(3)We design and implement a prototype system to assist the construction of target datasets.The system includes three modules: data processing and generation,dataset analysis and evaluation,and dataset auxiliary construction.It can provide support for self-built target datasets,evaluation target datasets,and optimization target datasets in practical engineering applications.