基于改进ResNet-50的图像特征提取网络

为了提高图像的特征质量,保证最后提取到的特征高度精炼,提出了一种新的方法；该方法首先将低分辨率图像经过小波变换分解成高频分量和低频分量,并结合插值法进行插值,最后通过小波逆变换得到高分辨率图像来为后续的特征提取提供高质量的图片输入；接着,选取ResNet-50网络作为基础网络,将Efficient Channel Attention(ECA)模块与ResNet残差结构结合形成一个全新的ECA-ResNet50模块,ECA模块具有的通道级的注意力机制,可以让整个网络更加专注于提取显著特征；经实验测试,该方法对于图像特征提取的质量有着明显的提升,均方误差下降可达6.65；结果表明,该方法可行有效,具有良好的工程应用前景;

Image feature extraction network based on improved ResNet-50

In order to improve the quality of image features and ensure that the final extracted features are highly refined, a new method is proposed;Firstly, the low resolution image is decomposed into high frequency component and low frequency component by wavelet transform, and interpolated by interpolation method;Finally, the high resolution image is obtained by inverse wavelet transform to provide high quality image input for subsequent feature extraction;Then, ResNet-50 network is selected as the basic network, and Efficient Channel Attention(ECA) module is combined with ResNet residual structure to form a new ECA-Resnet50 module; ECA module has a channel-level attention mechanism, which can make the whole network more focused on extracting salient features;Experimental results show that this method can significantly improve the quality of image feature extraction, and the mean square error can be reduced up to 6.65; The results show that the method is feasible and effective, and has a good engineering application prospect;