
Traditional underwater imaging suffers from issues, such as color deviation, low contrast, and blurred details, due to inconsistent attenuation of different light wavelengths and complex impurities in water. Currently, convolutional neural networks (CNNs) are one of the primary architectures for deep learning-based underwater image enhancement techniques. However, CNNs' limited global modeling capabilities result in enhanced images that still bear imperfections, leaving room for improvement. To push the boundaries of CNN performance in underwater image enhancement, an innovative network, namely, U-shift window Transformer (U-SWformer), which seamlessly integrates CNNs with Transformers, is proposed in this article. U-SWformer aims to vanquish the image degradation issues of traditional underwater imaging technology. Built upon the U-Net architecture, this network presents a shifting window self-attention module, enhancing the fusion of global image information. This fusion not only significantly boosts detail recovery and color correction but also elevates the overall visual fidelity. The U-SWformer network employs a multilayer loss integration strategy during training, including pixel-level loss in both RGB and lightness, a, b (LAB) color spaces to evaluate and guide color restoration, as well as perceptual loss and adversarial loss to evaluate and guide the reconstruction of high-quality image details. Experimental results underscore U-SWformer's remarkable advantages in image enhancement quality. It not only surpasses existing methods across various metrics but also achieves an optimal balance in parameter efficiency. These findings unequivocally validate the effectiveness and superiority of the U-SWformer network in revolutionizing underwater image enhancement.
This article presents the modeling and control of a bioinspired robotic fish with a three-segment tail, designed to achieve efficient and agile locomotion. A nonlinear model predictive control (MPC) system regulates the robotic fish’s forward speed, optimizing thrust in oscillating foils by maintaining the Strouhal number within ideal limits. This approach allows the robotic fish to maintain equilibrium states while operating within hydrodynamic conditions associated with efficient swimming. It ensures smooth velocity transitions even during dynamic speed changes, such as reaching a target velocity or pursuing a varying speed object, making it valuable in scenarios requiring both adaptability and high energy efficiency. The control system is based on the fish average behavior, isolating the fin-induced disturbances. An extended Kalman filter is included to ensure robust response under model uncertainties. The efficacy of the proposed approach is validated through simulations, where it is compared to a conventional proportional–integrative–derivative (PID) controller. Results show that the PID controller fails to guarantee efficiency in terms of the Strouhal number during speed transitions. In contrast, the nonlinear MPC successfully enforces the critical nonlinear constraint that is essential for optimal swimming dynamics.