CAJ | 학술논문

为了进一步提高超分辨图像重建效果，针对前期研究的超分辨误差反向传播神经网络（back propa-gation neural network，BPNN）重建结果中存在的块痕迹问题加以改进和优化。对影响 BPNN 超分辨效果的两个关键问题进行改进：（1）网络训练样本问题，将8×8→16×16的映射方式改进为2×2→4×4的映射方式，同时，采用相邻仅间隔一个像素的方式优化构造训练样本；（2）加速网络训练收敛问题，将网络训练规则由 BP 算法改进为改进的比例共轭梯度算法。网络训练实验和泛化实验表明，改进方法增加了网络训练样本数量，改善了超分辨BPNN 的输出图像质量，有效解决了超分辨结果中的块痕迹问题，使超分辨结果图像的峰值信噪比提高约8 dB。
위료진일보제고초분변도상중건효과，침대전기연구적초분변오차반향전파신경망락（back propa-gation neural network，BPNN）중건결과중존재적괴흔적문제가이개진화우화。대영향 BPNN 초분변효과적량개관건문제진행개진：（1）망락훈련양본문제，장8×8→16×16적영사방식개진위2×2→4×4적영사방식，동시，채용상린부간격일개상소적방식우화구조훈련양본；（2）가속망락훈련수렴문제，장망락훈련규칙유 BP 산법개진위개진적비례공액제도산법。망락훈련실험화범화실험표명，개진방법증가료망락훈련양본수량，개선료초분변BPNN 적수출도상질량，유효해결료초분변결과중적괴흔적문제，사초분변결과도상적봉치신조비제고약8 dB。
To solve the problem of block traces in the super-resolution results,an improved back propaga-tion (BP)neural network (BPNN)for super-resolution reconstruction (SRR)is established to further improve SRR image quality.Two important problems which directly affect super-resolution results are solved.First,the problem of BPNN training samples is solved.The mapping mode of 8×8→16×16 is improved as 2×2→4×4, at the same time,orders of training samples construction are optimized in a mode of one pixel interval.Second, the problem of speeding up net training convergence is solved.The net training rule is improved from BP algo-rithm to the improved scaled conjugate gradient algorithm.Experiment results show that the improved method increa-ses the quantity of training samples,enhances the SRR quality of BPNN output results images,and effectively solves the block traces problem of SRR results.The peak signal noise ratio of the SRR image increases about 8 dB.