OriginWater (China)

Purpose This study investigates the application of image captioning technology in automated bridge inspection. Given the scarcity of research in this domain, this study aims to evaluate the feasibility, effectiveness and practical implications of using transformer-based models to generate natural language descriptions of bridge damage from visual data. Design/methodology/approach A triangulated research methodology was used, comprising a systematic literature review to assess the current state of image captioning in bridge inspection; a feasibility study using an encoder–decoder architecture (EfficientNet–transformers) trained on a structural damage dataset; and an interview-based transferability study with seven industry professionals to evaluate practical adoption challenges. Findings The systematic review identified only four relevant studies, underscoring the nascent state of research in this field. The feasibility study demonstrated promising results, with EfficientNet–transformers achieving high bilingual evaluation understudy (BLEU) scores (BLEU-1:0.944, BLEU-4:0.904) in structural damage description tasks. Finally, industry feedback highlighted potential benefits in inspection efficiency but emphasized challenges in workflow integration and model reliability. Originality/value To the best of the authors; knowledge, this study represents one of the first comprehensive explorations of image captioning for bridge inspection, contributing both methodological and practical insights. It identifies key research gaps, including the need for domain-specific data sets and standardized evaluation frameworks, while proposing actionable directions for future AI applications in infrastructure maintenance. The findings provide a foundation for advancing automated inspection technologies toward safer and more efficient infrastructure management.

Human motion capture is crucial for various fields, but traditional optical systems (OMC) are costly and restrictive. Monocular video-based methods offer accessibility, yet face accuracy challenges, especially in dynamic sports like tennis. This study validates Capture4D, a novel Transformer-based monocular system, for capturing a wide range of tennis strokes. We developed a universal biomechanical analysis framework (K0-K5) applicable to twelve fundamental stroke types. To demonstrate the system's capabilities, this paper focused on a detailed validation using the tennis serve as a representative example. We conducted experiments with 9 high-level tennis players, and motion data were simultaneously captured using Capture4D (single RGB camera) and OMC Qualisys (gold standard). Accuracy was evaluated by comparing 3D joint coordinates and joint angles using Normalized Mean Per Joint Position Error (NMPJPE), RMSE, and MAE. The results demonstrated that Capture4D effectively captured the tennis player's motion, with average NMPJPE for tennis serves ranging from 69.5 mm to 88.3 mm, within the acceptable range (70-130 mm) for coaching purposes. Compared to OMC, Capture4D demonstrated comparable joint angle trajectories, with advantages in operational convenience, cost-effectiveness, and wider applicability. It offered an approximately 50% reduction in setup time and 80% cost savings. Capture4D presents a valid and practical monocular motion capture solution for coaching tennis and other broader applications in sports. While slightly less precise than OMC, its accuracy is acceptable for many use cases in coaching and teaching. It offers significant advantages in convenience and cost, paving the way for accessible motion analysis in diverse environments like outdoor settings and multi-person scenarios, in which OMC is not possible to be used. This technology holds promise for democratizing motion capture in sports training and coaching/teaching.

ABSTRACTVisible to short‐wave infrared multispectral imaging is gaining significant attention across various fields, including agriculture, security, and medical diagnostics. Traditional multispectral imaging systems often rely on separate sensors for different spectral bands, leading to complex optical alignment and irreversible resolution loss. Here, we present hardware‐algorithm co‐designed architecture to achieve multispectral super‐resolution imaging. Specifically, we demonstrate a monolithic quad‐spectral photovoltaic imaging platform featuring a resolution of 640 × 512 pixels with <1% dead pixels per channel. The system achieves broadband spectral integration from visible to short‐wave infrared (350–2350 nm) by combining an all‐polymer bulk heterojunction with colloidal quantum dots within a single CMOS‐compatible architecture. The compatibility of all‐polymer bulk heterojunction with direct photopatterning allows for precise patterning and high‐density integration, enabling the devices to operate efficiently in photovoltage mode. To address resolution degradation inherent in planar‐integrated spectral sensing architectures, we applied a super‐resolution reconstruction method, restoring images to a resolution of 640 × 512. The demonstrated capability to simultaneously capture and process multispectral data paves the way for CMOS integration, multispectral Imagers, organic photodetector, super‐resolution reconstruction applications in diverse fields, from precision agriculture to medical diagnostics and beyond.

In the process of analyzing the load characteristics of aggregation-type new business entities, linear regression methods are usually relied on to predict future changes in load characteristics. These methods only consider the linear relationship between load features and influencing factors, resulting in inaccurate prediction results in dynamic scenarios. Therefore, a new type of load characteristic prediction model for aggregated business entities based on machine learning is proposed. The spatial autoregressive model is applied to learn historical load data, obtain the load characteristic curves of aggregated new business entities, decompose the load characteristic data through the Prophet algorithm, calculate the Shapley value of each data component, measure its contribution to the load characteristic prediction results, and screen key data components to participate in subsequent prediction analysis. By integrating the improved feature pyramid network, multi-scale temporal convolutional network and Transformer network, a load characteristic prediction model based on machine learning is constructed to capture and screen the multi-scale features contained in the data components, thereby obtaining the predicted values of the load characteristics. The experimental results show that the relative error of the prediction results of this model is less than 1%, achieving an accurate prediction of the load characteristics of the new type of business entities of the aggregation type.

Generative recommendation has emerged as a transformative paradigm in recommender systems, enabling modeling user behavior autoregressively without explicit target conditioning. While this approach eliminates the need for target signals, it necessitates compressing extensive historical interactions-potentially spanning lifelong sequences-into coherent interest representations. Conventional methods for handling long sequences typically rely on target-guided search mechanisms (e.g., SIM) to efficiently filter and compress behaviors. However, this strategy is incompatible with generative frameworks due to their target-agnostic nature. To address these challenges, we propose a novel encoder-decoder model named HiCoGen (Hierarchical Compression-based Session-wise Generative Model), which efficiently models long-term interests in generative models. In the encoder, HiCoGen compresses behavior sequences using hierarchical content similarity clustering and employs a hierarchical attention architecture to reduce sequence length while preserving information integrity. In the decoder, HiCoGen uses session-wise generation instead of point-wise generation to better align with industrial short-video applications. To enhance the stability of session-wise generation, we introduce an auxiliary Hierarchical Multi-Token Prediction module. Extensive experiments on public and industrial datasets show significant performance gains over state-of-the-art methods (21.2% in ML-1M and 35.6% in industrial datasets on NDCG@3). We also conducted visualization and performance analysis to explore the advantages of long sequence modeling.