In the field of three-dimensional (3D) display technology [[1], [2], [3]], integral imaging is a light field reconstruction technique based on microlens arrays that captures and reproduces the direction and intensity of light rays in space [4,5]. This allows real 3D images to be viewed directly with the naked eye, without the need for special glasses [[6], [7], [8]]. Owing to its advantages of naked-eye viewing, continuous parallax, and full light field information recording, integral imaging is regarded as a key solution for next-generation true 3D displays [[9], [10], [11], [12]].Recent advancements in this field have further improved its resolution and viewing angle, facilitating applications in interactive and immersive displays [13,14].Moreover, novel approaches such as dual-view integral imaging and high-resolution displays have expanded the technology's applicability to various sectors, including medical imaging and augmented reality [15].

In integral imaging systems, the elemental image array (EIA) serves as the core data format, and its cross-device conversion is essential for ensuring compatibility across different display parameters and enabling broader applications [[16], [17], [18]]. However, differences in hardware parameters among various display terminals, such as microlens curvature radius, viewpoint sampling density, and lens array pitch, lead to significant discrepancies in the EIAs captured or generated for different devices [9,[19], [20], [21], [22]]. This parameter mismatch hinders the sharing and adaptation of integral imaging content across multiple platforms.

In integral imaging systems, EIAs are typically generated either by capturing real objects with a camera array or by synthesizing multi-view images based on 3D scene information, and their generation is usually constrained to specific display devices. Early work by Okano et al. (1997) introduced the direct recording method, using a large-aperture convex lens to capture parallel rays and applying image flipping to correct depth inversion, but it was limited to virtual-mode displays [23]. Martínez-Corral et al. (2005) proposed Smart Pixel Mapping (SPM), generating distortion-free, depth-correct real-image EIAs from multi-view captures, although the results remained dependent on the display device parameters [24,25]. Navarro et al. (2010) developed the SPOC (Smart Pseudoscopic-to-Orthoscopic Conversion) algorithm, which converts pseudoscopic EIA sources into orthoscopic ones to correctly represent the depth ordering of a 3D scene on a specific display [26,27]. In recent years, the rapid development of deep learning has led researchers to apply neural network techniques to the generation and enhancement of elemental image arrays (EIAs). For example, one study proposed a real-time optical reconstruction method based on path tracing and convolutional neural network (CNN) super-resolution, improving light field quality and EIA resolution [28]. Another study synthesized EIA from a single monocular image using forward light tracing and further refined the 3D details with a super-resolution network [29].Additionally, other studies have proposed various deep learning methods to improve EIA generation, such as optimizing monocular depth estimation using dense atrous pyramids and hybrid-scale UNet, addressing gradient vanishing issues in depth estimation with the URNet model, and supporting EIA generation in dynamic scenes through unsupervised learning with optical flow features and multiple constraints [[30], [31], [32], [33]].Despite these advances, challenges remain in adapting EIAs across different display devices. For instance, an EIA generated under specific device parameters—such as those of a smartphone display—often cannot be correctly displayed on a tablet due to differences in device parameters.

To address the aforementioned incompatibility issue, we propose an algorithm capable of converting existing EIA content generated for one device into EIA data adapted to other devices with different optical parameters. The method first extracts multi-view images from the source EIA, which are then processed by a convolutional residual neural network to generate multi-view images adapted to the target device. The final EIA is re-encoded according to the target device's parameters. This approach does not rely on additional capture and enhances cross-device adaptability of EIA data among heterogeneous devices.

In this study, the feasibility of the proposed method was evaluated using a light field dataset comprising diverse typical scenes and combinations of source and target devices with different optical parameters. The experimental results show that this method can convert elemental image array content from the source device to the target device, enhances cross-device compatibility, and enables more flexible utilization of integral imaging content across devices with varying optical configurations.