Research On Image Fusion And Colorization Technology Based On Transfer Learning

Image quality enhancement can enhance the interpretability of the image,improve the quality and definition and then increase the value of the image.Multi-focus image fusion and image colorization,as the typical algorithms of image quality enhancement technology,have always been image quality enhancement and Research hotspots in the field of image processing.Multi-focus image fusion is to merge two or more(usually two)images containing different focus areas in the same scene into one image,so that the fused image presents a fully focused scene.Image colorization technology is a technology that converts grayscale images or single-channel images into color natural light images,which can significantly improve the visual effect of the image and enhance the key features of the image.The above two methods can significantly enhance the quality of the image,so that the processed image contains richer information,which is convenient for the machine to process the image.In recent years,many scholars have applied deep learning to this field,but training a well-performing deep neural network model usually requires a large amount of labeled data for model training,and collecting and labeling a large amount of training data often It takes a lot of manpower and material resources,and the use of massive data sets to train the model is time-consuming.In order to solve the above problems,based on the theory of deep learning and transfer learning,this paper uses neural networks based on transfer learning to realize the fusion of multi-focus images and the colorization of gray-scale remote sensing images.In the work of multi-focus image fusion based on transfer learning,this paper proposes a transfer learning network model containing multi-level hybird loss functions.First,use the VGG-19 network to perform preliminary feature extraction on the inputted pair of multi-focus source images,and then transfer the preliminary extracted features to a multi-level hybird loss function,a multi-level skip-connection structure,multiple encoders and decoders.In the neural network;through five different encoders to perform deep feature extraction and dimensionality reduction on the source image;and then input the extracted features into the corresponding five different decoders to reconstruct the feature map,and finally get used preliminary decision map for pixellevel multi-focus image fusion.Then use the post-processing method to optimize the preliminary decision map to obtain the secondary decision map,and finally use the secondary decision map to fuse the source image to obtain a clear image with full focus.By comparing with other methods,the all-focus image generated by this method has excellent performance in both subjective visual evaluation and objective index evaluation.In the colorization of remote sensing images,this paper proposes a deep learning network including multi-scale convolution and multi-level attention mechanism based on transfer learning to generate color remote sensing images.First,we use the VGG-19 network to perform preliminary feature extraction on the grayscale remote sensing image,and then input the features extracted by the VGG-19 network into the multiscale convolution module,then perform multi-scale feature extraction and dimensionality reduction on these features,and then input the features to the decoder network for feature decoding and reconstruction,and fuse the feature map output by the multi-scale convolution module through the SE attention mechanism module and the feature map output by the decoder.Through the layer decoding and feature reconstruction of the decoder,finally generate color remote sensing images.By comparing with other methods,the method proposed in this paper has a good visual effect on the colorization of remote sensing images,and is better than other methods in terms of image quality indicators.