Jinyang Li

Imaging systems work diversely in the image processing domain, and each system contains specific characteristics. We are developing models to fuse images from different sensors and environments to get promising outcomes for different computer vision applications. The multiple unified models have been developed for multiple tasks such as multi-focus (MF), multi-exposure (ME), and multi-modal (MM) image fusion. The careful tuning of such models is required to get optimal results, which are still not applicable to diverse applications. We propose an automatic machine learning (AML) based multi-tasking image fusion approach to overcome this problem. Initially, we evaluate source images with AML and feed them to the task-based models. Then, the source images are fused with the pre-trained and fine-tuned models. The experimental results authenticate the consequences of our proposed approach compared to generic approaches.

Images are obtained by perceiving the sunlight reflected by objects or scenes, and due to limited solar irradiance, spatial resolution certainly decreases. In contrast, multispectral sensors hold the spatial information of ground objects even though the obtained images only have a few bands, which means that the obtained images have high spatial but insufficient spectral resolution. To overcome these limitations, we have developed a spectral and polarization image fusion technique that improves several perspectives of image fusion, such as intensity, edges, textures, and substantial information. Our proposed method utilizes an encoder-decoder approach, which contains dense and convolutional blocks to simulate the source images. The polarization images have been used to develop DoLP images that help to improve fused images’ quality. We can perceive the visual quality of fused images from the experimental results.

The ultimate image with important information and better visuals can be obtained by fusing infrared and visual images. Several fusion methods consist of complex networks to derive the finest parameters from different source images and do not exploit reliable and salient information completely. To avoid the limitations of information loss and fusion of feature information, we have proposed a multilayer triple dense network for infrared and visible image fusion named as MTDFusion, which completely exploits salient features and utilizes the residual features from the combination of input images adaptively. The encoding network is designed as a triple network of dense blocks that takes inputs as infrared, visible, and the difference extracted from infrared and visible images. The fusion layer exploits the weighted combination of input images, and decoding layers work conventionally. The results obtained by our proposed approach reveal optimal image details and higher contrast. We have improved the loss function by combining smooth-ℒ ₁ and structured similarity index (SSIM) loss functions for our proposed method. Additionally, the different variations of λs in SSIM are examined using qualitative and quantitative assessments. The results achieve optimal fusion performance compared to several state-of-the-art methods. The source code for MTDFusion is available at https://github.com/tgg-77/MTDFusion.

Sparse representation has been proven to be a very effective technique for various image restoration applications. In this paper, an improved sparse representation based method is proposed to detect and estimate defocus blur of imaging sensors. Considering the fact that the patterns usually vary remarkably across different images or different patches in a single image, it is unstable and time-consuming for sparse representation over an over-complete dictionary. We propose an adaptive domain selection scheme to prelearn a set of compact dictionaries and adaptively select the optimal dictionary to each image patch. Then, with nonlocal structure similarity, the proposed method learns nonzero-mean coefficients’ distributions that are much more closer to the real ones. More accurate sparse coefficients can be obtained and further improve the performance of results. Experimental results validate that the proposed method outperforms existing defocus blur estimation approaches, both qualitatively and quantitatively.