多视角自适应融合的叶片光场图像超分辨方法OA

A light field image super-resolution method called LFAFormer(light field angular transformer)based on angular adaptive attention mechanism and multi-layer perception structure is proposed to improve the detail restoration ability in aerial blade imaging.On the basis of the convolutional neural network(CNN)framework,this method designs a multi-layer perception module to extract feature information at different semantic levels layer by layer,and introduces the perspective adaptive weight mechanism to dynamically weight and fuse different sub-views by means of angular attention,so as to improve the utilization efficiency of key angular information.To enhance the structural consistency among multi-views,the angular consistency loss function is introduced to constrain the reconstruction image to maintain angular coherence in the case of the angular transformation.Quantitative evaluation and comparative verification were carried out on multiple light field data sets(including the actual collected aerial blade images).The results show that the proposed method is superior to the current mainstream methods in terms of peak signal-to-noise ratio(PSNR),structural SIMilarity(SSIM)and edge root mean square error(EdgeRMSE),especially in terms of edge sharpness and detail recovery.This method can effectively improve the image quality and diagnostic accuracy of light field imaging in the field of aviation parts detection,and provide robust support for high-precision visual inspection system,so it has a good application prospect.