We present a sparse reconstruction light field (SRLF) framework using compressed sensing theory. Light field signals have sparsity in some specific phenomena, such as nonocclusions, smooth surfaces and texture uniformity. We study the light field sparsity to analyze the sparse basis and measurement matrix of the light field in a complicated scene. This extends previous work on light field sampling that considered either spatial or angular dimensions, which can be used to control the sampling rate of the light field. Furthermore, the sparse Bayes learning (SBL) algorithm is applied to the reconstruction of sparsely sampled light fields. We derive a learning machine for the light field SBL algorithm, which can improve the rendering quality based on a given set of captured multiview images. The proposed SRLF compares favorably with state-of-the-art light field sampling and reconstruction techniques. The innovation of the SRLF is to use compressed sensing theory to further reduce the light field sampling rate. We conduct a detailed derivation of the mathematical theory of light field sparseness.

Retinal blood vessels segmentation acts as an important part to the treatment of ocular disease. Lately, automatic segmentation based on deep learning can solves problems of low efficiency and strong subjectivity of manual segmentation, and attracts the attention of researchers. In this paper, a novel segmentation model called Residualpath-Res-Dense Net (RRD-Net) is proposed to achieve vessel segmentation. In RRD-Net, Res-block and Dense-block are used to help speeding up the convergence of the network and learning more intrinsic features. In addition, the introduction of Residual Path can cut down the semantic difference between the connected feature and eliminate the potential impact of semantic difference on segmentation accuracy. We apply benchmark datasets DRIVE and CHASE_DB1 to evaluate effectiveness of the proposed network. Accuracy, sensitivity and F1 score demonstrate the effectiveness of RRD-Net.

In the neuron morphological reconstruction aspect, denoting the type of node was a challenge by defining the radius of reconstructed nodes. In our research, we adapted deep learning methods to this task. The foreground pixel and background pixel of neuron images have a large imbalance in numbers, so that the model was hard to extract the features of neuron images, as a result, missing the neuron cell body during segmentation. In order to resolve this problem, we introduced the self-supervision signal on the top-level feature map by knowledge distillation using a variational autoencoder, except for changing the segmentation loss function. In order to resolve the flaw of hollows in neuron cell bodies, we introduce the attention fuse module on the skip connection from the segmentation encoder and upsampling layer of the segmentation decoder. Compared with the original model, the accuracy and Dice coefficient of the model in this paper are improved by 2% and 15%, respectively. As far as the two metrics of IOU and Dice coefficient are concerned, our method basically achieves the same level as the SwimUNet model, and there is no edge squaring in the segmentation results.