Underwater images often face visual degradation problems such as color deviations, low contrast, and blurred details. Previous studies suffer from poor visual perception performance and over-enhancement, which poses a challenge to improve the visibility of underwater images comprehensively and consistently. This paper proposes a framework for underwater image enhancement based on non-physical transformation and unsupervised learning (NPT-UL). The framework includes an image data augmentation strategy and an unsupervised learning model. Firstly, a data augmentation strategy based on non-physical transformation is proposed, which consists of three types of transformation processes to optimize the color deviation, contrast, and detail blur in the initial image data set. Secondly, an unsupervised learning model based on contrast learning and generative adversarial network is proposed to enhance the input underwater image. Three loss functions including adversarial loss, PatchNCE loss, and identity loss are consolidated into the network to ensure structural similarity and color authenticity. Finally, qualitative, quantitative, ablative experiments are presented to evaluate the performance using different underwater image datasets. Compared with the state-of-art methods, NPT-UL has better performance which can effectively improve the image visual quality. Tested in the UIEB data set, the values of UCIQE, CCF, UIQM, BC(e) and BC(r) metrics obtained by NPT-UL reach 0.644, 42.311, 4.325, 4.761 and 2.719, separately. The proposed framework can help to solve the problems of color distortion and feature loss, showing significant application potential in the field of underwater image restoration and target detection.

A biomimetic system using polymer lenses for the optical design and application is developed. The system mainly consisted of a bionic cornea lens, voice coil motor, compression ring, bionic crystalline lens, substrate, and CCD sensor. By controlling the current of the voice coil motor, we could change the motion of the compression ring to alter the curvature radius of the bionic crystalline lens, thus adjusting the focal length of the whole system. The integrated constructure of the optical system was presented, as well as the detailed description of the lens composition, material, and fabrication process. Images under different displacement loads were captured, the relationship among the curvature radius, observed back focal length, and predicted effective focal length was analyzed, and the spot diagram of the optical system was simulated using ZEMAX software. The focal length of the optical system ranged from 17.3 to 24.5 mm under a tiny displacement load from 0 to 0.14 mm. Besides, the images captured at different rotating angles presented almost identical patterns and the same image quality, which showed good robustness to the gravity. The biomimetic optical system is of interest to develop an integrated, low-cost, and stable imaging system.

A bionic optical imaging system with an aspheric solid–liquid mixed variable-focus lens was designed and fabricated. The entire system mainly consisted of a doublet lens, a solid–liquid mixed variable-focus lens, a connecting part, and a CCD imaging device. To mimic the structure of the crystalline lens, the solid–liquid mixed variable-focus lens consisted of a polydimethylsiloxane (PDMS) lens, a polymethyl methacrylate lens, and the liquid of ethyl silicone oil. By pumping liquid in or out of the cavity using a microinjector, the curvatures of the front and rear surfaces of the PDMS lens were varied, resulting in a change of focal length. The overall structure of the system was presented, as well as a detailed description of the solid–liquid mixed variable-focus lens, material, and fabrication process. Under different injection volumes, the deformation of the PDMS lens was measured and simulated, pictures were captured, and the optical performance was analyzed in simulations and experiments. The focal length of the system ranged from 25.05 to 14.61 mm, and the variation of the diopter was 28.5D, which was larger than that of the human eye.

Surface defect segmentation is the basis of industrial production, where pixel-level defect prediction plays a vital role in many applications. As a new trend, few-shot defect detection aims to recognize defects by combining information from a few defect samples. However, many traditional few-shot methods do not consider patch-level correlations between input images and memory banks. This work introduces a novel representation framework consisting hierarchical normal and defect memory banks for few-shot surface defect detection. We propose an adaptive feature bank updating scheme that dynamically integrates frequent features while discarding rare ones. We design a fine-grain feature replacement module rooted in a matching attention mechanism to highlight ambiguous regions for input images, and use fused features to identify overall defects. On benchmark data sets, our strategy can obtain in-depth insights from multiple scales, which showcase superiority in generalization for unidentified anomalies better than others in experiments. The experimental results for the 10-shot scenarios demonstrate that the proposed method achieves an average pixel AUC of 88.6%, 93.7%, and 98.7% on three benchmark data sets, respectively, and with more genuine and pseudo anomalous samples, the proposed approach can achieve 98.5% and 95.5% in image AUC and pixel F1 score respectively, outperforming the state-of-the-art with 1.2% and 24% in each metric.