Remote sensing ship recognition technology is one of the important research areas for achieving ocean security monitoring. However, maritime ship targets features such as multi-scale, arbitrary direction, and dense arrangement, and their imaging is often influenced by factors such as sea fog and sea surface background. It is challenging for fast and accurate detection of remote sensing ships. We propose a new ship detection algorithm YOLO-Remote Sensing Ship Detection (YOLO-RSSD) based on YOLOv5, which effectively improves the accuracy of ship detection while ensuring detection speed. It is attributed to our proposed four optimization measures in this model: First, the K-means++ algorithm is introduced in the data preprocessing part to improve the clustering effect and obtain candidate box sizes suitable for multi-scale ship dataset. Then, an improved BiFPN structure is embeded in the feature fusion part to achieve cross-layer multi-scale weighted feature fusion. Thirdly, the original bounding box regression loss function is replaced with the EIoU loss, which is effective for accelerating the convergence and improving the regression accuracy of predicted boxes. Finally, a channel attention mechanism is introduced in the convolutional unit to enhance the model's ability to capture ship features. Experimental results show that the YOLO-RSSD model achieves a detection accuracy of 96.1 $\%$ for remote sensing ships, which is 4.3 $\%$ higher than the original YOLOv5 network. In addition, YOLO-RSSD performs good robustness and generalization ability. This means that our method has high practical value and provides a new solution for the analysis and application of remote sensing images.

An improved deep neural network incorporating attention mechanism and DSSIM loss function (AM_U_Net) is used to recover input images with speckles transmitted through a multimode fiber (MMF). The network is trained on a relatively small dataset and demonstrates an optimal reconstruction ability and generalization ability. Furthermore, a bimodal fusion method is developed based on S polarization and P polarization speckles, greatly improving the recognition accuracy. These findings prove that AM_U_Net has remarkable capabilities for information recovery and transfer learning and good tolerance and robustness under different MMF transmission conditions, indicating its significant application potential in medical imaging and secure communication.

A hybrid method of combining Neural Networks (NNs) and optimization algorithms is proposed for the inverse design of orbital angular momentum (OAM) transmission fiber with high efficiency and precision. NNs are used to predict the optical properties of OAM transmission fibers with high calculation accuracy and speed, including chromatic dispersion and effective index difference ( $\Delta n_{eff}$ ). Then the trained prediction models are combined with particle swarm optimization (PSO) and multi-objective particle swarm optimization (MOPSO) algorithms for the inverse design of OAM transmission fiber respectively. After analyzing the differences and properties of the hybrid PSO-NN and MOPSO-NN algorithms in detail, we designed an OAM transmission fiber with extreme low chromatic dispersion through the hybrid MOPSO-NN algorithm. The proposed method can avoid the single solution problem of tandem neural network and output a number of appropriate fiber structures according to the design requirements, which provides a new approach for the inverse design of optical structure with extreme performance. Finally, the accuracy and effectiveness of the proposed hybrid method are compared and demonstrated through COMSOL Multiphysics.

Due to the occurrence of redundant speckle, multimode fiber (MMF) imaging is extremely challenging. Our work studies the relationship between the effective feature distribution of the speckle field and the local spatial position and area, and proves that the information distribution of the speckle is highly redundant. The effective feature refers to the phase and amplitude information of the optical field carrying the image point information and the co-exciting very redundant information due to mode dispersion, interference, coupling, and entrained noise through transmission. The neural network Swin-Unet can well learn the association information between global and local features, greatly simplifies the fitting of the MMF end-to-end global mapping relationship, and achieves high-fidelity reconstruction from the local speckle field to the global image. This work will contribute to the realization of MMF real-time large-field endoscopic imaging.