Suitable Matching Area Selection Method Based On Deep Learning

In the terminal guidance phase of missiles,scene matching technology is often used as a navigation means.The selection of scene matching area is the key factor affecting the performance of scene matching system,and it is the core technology in the preparation of reference map.Therefore,it is of great practical significance to establish a adaptable selection criteria for suitable matching area.The existing scene area matching suitability analysis work,whether based on optical images or SAR images,extracts the traditional image feature parameters related to region matching suitability,and further selects the matching area layer by layer according to the threshold value,or a classifier is thus trained to make predictions about the performance of suitable matching.Based on the previous work,this thesis studies the matching suitability analysis method of scene matching,and proposes a ResNet-based technique used for suitable matching area selection.The main research contents are as follows:For the evaluation ambiguity of current scene area matching suitability researches,the matching probability index is used to quantify the suitable matching performance.Combining with the engineering application,the thesis investigates the matching performance of multiple reference areas after imaging in different scenarios.For a reference picture,a plurality of sub-pictures of real-time picture size are taken and affine transformed,and a large number of simulated real-time images are generated.The simulated real-time images are matched with the reference area,and the number of real-time images correctly matched is counted and the probability is calculated as a quantitative index to evaluate the suitable matching performance of the current reference area.Aiming at the existing difficulty of designing the image parameters related tosuitable matching performance,the deep convolutional neural network method is used to extract high-dimensional image features.Combining feature extraction with regression training,this method use massive reference areas of typical size as the sample inputs,and the corresponding matching probability values obtained by the mean-normalization algorithm in these regions as sample outputs,and the ResNet network is trained as matching probability prediction model.The implementation quickly gives an relatively accurate matching probability prediction for the specified reference area.Finally,statistical experiments were carried out to analyze the matching suitability of the reference areas from different feature categories.The scene area matching suitability analysis method proposed in the thesis is compared with the traditional statistical feature-based suitable matching area selection method.The experimental results show that the thesis method has higher classification accuracy under the condition of ensuring high real-time performance,and it has higher flexibility for the selection of the suitable matching area.