基于可感知网络模型的多尺度特征图像重构研究

Research on Multi-scale Feature Image Reconstruction Based on Perceptible Network Model

为重点解决深度重构模型多尺度特征表示不足和自适应性处理能力不强问题,本研究设计了压缩感知多尺度图像重构深度网络模型。以引入深度卷积神经网络模型提取图像特征为前提,利用空间金字塔池化提高多尺度特征描述能力,设计噪声滤波器强化噪声控制,改进非局部神经网络结构以提高模型的自适应性能。对比实验结果表明,本研究设计的图像深度重构模块,压缩编码后解码重构图像的方差、变异系数和信噪比控制良好;图像相位一致性特征和梯度特征质量评价较高;在采样率为50%~60%时,峰值信噪比值最高;初始图像和解码重构图像之间的相对范数l_(2)误差在0.16左右,峰值信噪比大于46dB,图像余弦相似度评价的余弦值为0.91左右、余弦夹角值为0.13°左右,特征相似性为0.80左右,结构相似比值为0.86左右,说明该模型提取的图像多尺度特征准确性和精细性较高,深度重构模块自适应性较强,误差控制准确,组成模块之间实现了高度融合。

The compressed sensing multi-scale image reconstruction depth network model designed in this study focuses on solving the problems of insufficient multi-scale features representation and weak adaptive processing ability of the depth reconstruction model.On the premise of introducing a deep convolutional neural network model to extract image features,the spatial pyramid pooling was utilized to improve multi-scale feature description ability,the noise filters were designed to strengthen noise control,and the non-local neural network structure was improved to enhance the adaptive performance of the model.The experimental results showed that the designed image depth reconstruction module has good control over variance,coefficient of variation,and signal-to-noise ratio of the decoded reconstructed image after compression and encoding.The quality evaluation of image phase consistency and gradient features is high.When the sampling rate is between 50%and 60%,the peak signal to noise ratio value is the highest.The relative norm l_(2) error between the initial image and the decoded reconstructed image is around 0.16,and the peak signal to noise ratio is greater than 46dB,The cosine value of image cosine similarity evaluation is about 0.91,the cosine angle value is about 0.13°,the feature similarity is about 0.80,and the structural similarity ratio is about 0.86.The model designed in this study has strong adaptability,accurate error control,and achieves high fusion between the constituent modules.