Research On Image And Text Relevance Methods In Cross-media Information For Aviation Safety Incidents

With the constant development of people’s living standards,more and more people have begun to choose to travel through the civil aviation system.At the same time that the scale of civil aviation flights has grown,some hidden dangers that may affect the course of civil aviation flights have gradually surfaced.Aviation safety incidents are suddenly incidents related to the aircraft operation that cause individuals injury or death,economic loss,aircraft damage or disappearance.At the same time,the explosive growth of information resources and the rapid spread of network information have also created great challenges for data collection and collation of aviation safety incidents.Therefore,it is of great significance to study the image and text relevance method for cross-media information and apply it to the knowledge construction of aviation safety incidents.This article has studied the following contents.In order to solve the question that an aviation safety incidents modal simplifical,a method based on recurrent residual network is proposed for image and text retrieval.The main framework of this method is a two-branch embedded network containing identity connections and recurrent connections,which reduces the amount of calculations while extracting deeper features.The improvement of the algorithm mainly in terms of the use of a Hybrid Gaussian-Laplacian Mixture Model and Gaussian Mixture Model used in the text feature extraction step,and the objective function uses a two-way ranking loss based on triples.We employ R@K as experimental indicator.Experiments prove that this setting has some improvement on the experimental effect.On the public dataset,the R@K value of the two-way retrieval compared with the orignally recurrent residual network has increased by nearly 1%.When the feature embedding method applied to the images and text bi-directional retrieval of aviation safety incidents,the experimental results was better than the network without embedded components.In order to achieve more accurate images and text matching for aviation safetyincidents,a two-way CNN model is used,which is also a dual branch like the retrieval network.Residual connections and convolutional layers are used to extract features on the embedded branch,and add a end-to-end instance loss is fine-tuned.Finally,an objective function is designed based on the cosine distance,combining the instance loss and the ranking loss.The model is applied to the aviation safety incidents dataset,and the results are compared with the previous related methods,which proves that the method effectively improves the recall rate of about 7 percentage points on images and text matching results.Furthermore,which provides effective method support for the information modal expansion and information collation of civil aviation field.