Research On The Image Intelligence Target Detection And Scene Recognition Technology Based On Convolutional Neural Network

In recent years,with the increasing degree of informatization of battlefields,the amount of image intelligence collected by various sensors has been increasing,and image intelligence target detection and scene recognition have become a research hotspot in the military field.For the massive image intelligence data,using computer to automatically extract the target and scene elements of image intelligence and transform it into structured or semi-structured data is an important step in battlefield intelligence data mining.It is of great significance for helping commanders to make decision and mastering the battlefield initiative.According to the above actual requirments,this paper draws on the convolutional neural network framework widely used in the field of image recognition.Based on the specific characteristics of military images,this paper studies the military image target detection and scene recognition methods based on convolutional neural networks.Including the following aspects:1.A military image target detection framework based on Mask R-CNN is designed.This paper analyzes and contrasts the military image target detection task and the common natural image target detection task,and applies the current framework Mask R-CNN in the target detection field to the military image target detection task.The improvement is made in the Anchor ratio,target area threshold and confidence threshold of Mask R-CNN framework.2.The training method based on layer freeze migration learning is designed to train the Mask R-CNN network.Based on the current situation of military image target detection training data lacking,combined with migration learning theory,this paper designs a military image target detection framework based on Mask R-CNN and migration learning,and uses the layer freezing method to train the Mask R-CNN network to realize Mask R-CNN network training on small sample data sets,the effectiveness of the method is verified by experiments.3.A military image scene detection method based on CNN and semantic information is proposed.Firstly,combined with the specific characteristics of military images,the classic AlexNet network is improved as a feature extraction network.On this basis,this paper proposes the concept of correction coefficient for the semantic relationship between the target and the scene in the military image.The feature extraction and initial classification are performed by convolutional neural network,the initial classification probability is output,and the initial classification probability is corrected by the correction coefficient,which reduce the probability of misclassification when using the convolutional neural network for scene recognition,and effectively improve the military image scene recognition accuracy.4.The image intelligence target scene feature extraction software tool prototype system is designed,developed and verified.The software tool prototype integrates the target detection and scene recognition algorithms proposed in this paper,and develops visual display and result storage functions.This paper verifies the software tool prototype using military images in a typical natural scene.The results show that the software tool prototype can effectively extract the image intelligence target scene elements,and verify the effectiveness of the method and the availability of the tool.