Inference Time Research Articles

Application of artificial intelligence in VSD prenatal diagnosis from fetal heart ultrasound images.

Developing a combined artificial intelligence (AI) and ultrasound imaging to provide an accurate, objective, and efficient adjunctive diagnostic approach for fetal heart ventricular septal defects (VSD). 1,451 fetal heart ultrasound images from 500 pregnant women were comprehensively analyzed between January 2016 and June 2022. The fetal heart region was manually labeled and the presence of VSD was discriminated by experts. The principle of five-fold cross-validation was followed in the training set to develop the AI model to assist in the diagnosis of VSD. The model was evaluated in the test set using metrics such as mAP@0.5, precision, recall, and F1 score. The diagnostic accuracy and inference time were also compared with junior doctors, intermediate doctors, and senior doctors. The mAP@0.5, precision, recall, and F1 scores for the AI model diagnosis of VSD were 0.926, 0.879, 0.873, and 0.88, respectively. The accuracy of junior doctors and intermediate doctors improved by 6.7% and 2.8%, respectively, with the assistance of this system. This study reports an AI-assisted diagnostic method for VSD that has a high agreement with manual recognition. It also has a low number of parameters and computational complexity, which can also improve the diagnostic accuracy and speed of some physicians for VSD.

Decoding Continuous Tracking Eye Movements from Cortical Spiking Activity.

Eye movements are the primary way primates interact with the world. Understanding how the brain controls the eyes is therefore crucial for improving human health and designing visual rehabilitation devices. However, brain activity is challenging to decipher. Here, we leveraged machine learning algorithms to reconstruct tracking eye movements from high-resolution neuronal recordings. We found that continuous eye position could be decoded with high accuracy using spiking data from only a few dozen cortical neurons. We tested eight decoders and found that neural network models yielded the highest decoding accuracy. Simpler models performed well above chance with a substantial reduction in training time. We measured the impact of data quantity (e.g. number of neurons) and data format (e.g. bin width) on training time, inference time, and generalizability. Training models with more input data improved performance, as expected, but the format of the behavioral output was critical for emphasizing or omitting specific oculomotor events. Our results provide the first demonstration, to our knowledge, of continuously decoded eye movements across a large field of view. Our comprehensive investigation of predictive power and computational efficiency for common decoder architectures provides a much-needed foundation for future work on real-time gaze-tracking devices.

Camera-view supervision for bird's-eye-view semantic segmentation

Bird's-eye-view Semantic Segmentation (BEVSS) is a powerful and crucial component of planning and control systems in many autonomous vehicles. Current methods rely on end-to-end learning to train models, leading to indirectly supervised and inaccurate camera-to-BEV projections. We propose a novel method of supervising feature extraction with camera-view depth and segmentation information, which improves the quality of feature extraction and projection in the BEVSS pipeline. Our model, evaluated on the nuScenes dataset, shows a 3.8% improvement in Intersection-over-Union (IoU) for vehicle segmentation and a 30-fold reduction in depth error compared to baselines, while maintaining competitive inference times of 32 FPS. This method offers more accurate and reliable BEVSS for real-time autonomous driving systems. The codes and implementation details and code can be found at https://github.com/bluffish/sucam.

Using an improved U-Net++ with a T-Max-Avg-Pooling layer as a rapid approach for concrete crack detection

The monitoring of concrete structures has advanced remarkably with the aid of deep learning technologies. Since concrete is multi-purpose and low-cost, it is extensively used for construction purposes. Concrete is very enduring. Nevertheless, it tends to crack which endangers the integrity of the structure and results in complications. The current study offers a new image segmentation approach for detecting cracks in concrete by making use of an optimized U-Net++ architecture. The proposed model gives the features of the T-Max-Avg Pooling layer which effectively combines the advantages of traditional max and average pooling using a learnable parameter to balance feature extraction dynamically. This innovation both improves the output accuracy and processing speed and captures the fine details. In addition, it mitigates noise and transcends the limitations of conventional pooling methods. Moreover, using learnable pruning and shortening skip connections in U-Net++ reduce redundant computations, making the model faster without compromising accuracy. In comparison with other models like Mask R-CNN and VGG-U-Net, the proposed model had considerably faster inference times (21.01 ms per image) and fewer computational requirements (40G FLOPs), making it very suitable for real-time monitoring applications. The DeepCrack and Concrete Pavement Crack datasets were employed to assess the model thoroughly which yielded an MIoU score of 82.1%, an F1 score of 90.12%, a Dice loss score of 93.7%, and an overall accuracy of 97.65%. According to the results, the enhanced U-Net++ with T-Max-Avg Pooling provided a balanced trade-off between segmentation accuracy and computational efficiency. This indicates its considerable potential for automated real-time crack detection in concrete structures by employing resource-constrained environments including drones and mobile platforms.

Application of machine learning for detecting and tracking turbulent structures in plasma fusion devices using ultra fast imaging.

This study explores the application of machine learning techniques for detecting and tracking plasma filaments around the boundary of magnetically confined tokamak plasmas. Plasma filaments, also called blobs, are responsible for enhanced turbulent transport across magnetic field lines, and their accurate characterization is crucial for optimizing the performance of magnetic fusion devices. We present a novel approach that combines machine learning methods applied to data obtained from ultra-fast cameras, including YOLO (You Only Look Once) for object detection, semantic segmentation, and specific tracking methods. This approach enables fast and accurate detection and tracking of filaments while overcoming the limitations of conventional methods, which are time-consuming and prone to human subjectivity. A significant advance in our study lies in the development of a method for automatically labeling a large batch of data, which greatly facilitates the training of supervised machine learning algorithms. Using these techniques, we obtained promising results demonstrating a significant improvement over conventional tracking methods, achieving a detection accuracy of up to 98.8%, while reducing the inference time per frame by 15% to 31% compared to conventional Kalman filter tracking. These results open up new perspectives for investigating turbulent phenomena in tokamaks, and could have important implications for the development of controlled nuclear fusion.

A Reparameterization Feature Redundancy Extract Network for Unmanned Aerial Vehicles Detection

In unmanned aerial vehicles (UAVs) detection, challenges such as occlusion, complex backgrounds, motion blur, and inference time often lead to false detections and missed detections. General object detection frameworks encounter difficulties in adequately tackling these challenges, leading to substantial information loss during network downsampling, inadequate feature fusion, and being unable to meet real-time requirements. In this paper, we propose a Real-Time Small Object Detection YOLO (RTSOD-YOLO) model to tackle the various challenges faced in UAVs detection. We further enhance the adaptive nature of the Adown module by incorporating an adaptive spatial attention mechanism. This mechanism processes the downsampled feature maps, enabling the model to better focus on key regions. Secondly, to address the issue of insufficient feature fusion, we employ combined serial and parallel triple feature encoding (TFE). This approach fuses scale-sequence features from both shallow features and twice-encoded features, resulting in a new small-scale object detection layer. While enhancing the global context awareness of the existing detection layers, this also enriches the small-scale object detection layer with detailed information. Since rich redundant features often ensure a comprehensive understanding of the input, which is a key characteristic of deep neural networks, we propose a more efficient redundant feature generation module. This module generates more feature maps with fewer parameters. Additionally, we introduce reparameterization techniques to compensate for potential feature loss while further improving the model’s inference speed. Experimental results demonstrate that our proposed RTSOD-YOLO achieves superior detection performance, with mAP50/mAP50:95 reaching 97.3%/51.7%, which represents improvement of 3%/3.5% over YOLOv8, and 2.6%/0.1% higher than YOLOv10. Additionally, it has the lowest parameter count and FLOPs, making it highly efficient in terms of computational resources.

TIA-UNet: transformer-enhanced deep learning for adolescent idiopathic scoliosis spinal x-ray image segmentation

Abstract Adolescent Idiopathic Scoliosis (AIS) is a common spinal deformity where precise diagnosis is crucial for developing effective treatment strategies. Traditional manual x-ray image analysis is time-consuming and highly dependent on the operator’s expertise, thus constraining diagnostic efficiency and accuracy. This study aimed to develop an automated thoracolumbar spine segmentation method utilizing deep learning to enhance the efficiency and accuracy of AIS diagnosis. We introduced TIA-UNet, an innovative network architecture that combines Convolutional Neural Networks (CNN) with Transformer models. By integrating the IR Block and DA Block, TIA-UNet was optimized for feature extraction and multi-scale information fusion. The model underwent training and validation using our established Adolescent Scoliosis Medical Dataset (ASMD). TIA-UNet attained a Dice similarity coefficient of 90. 02%, a Mean Intersection over Union (MIoU) of 81. 96%, and a Hausdorff Distance (HD) of 4. 09, significantly surpassing current state-of-the-art methods such as UNet and TransUNet. Moreover, TIA-UNet exhibited superior computational efficiency regarding parameter count, inference time, and floating-point operations (FLOPs). As an automated medical image segmentation algorithm, TIA-UNet enhanced segmentation accuracy while maintaining high computational efficiency, demonstrating significant potential for clinical diagnostic applications. This study provides compelling evidence supporting the utilization of deep learning techniques in medical image analysis.

YOLOv8-seg-CP: a lightweight instance segmentation algorithm for chip pad based on improved YOLOv8-seg model.

Real-time detection and accurate segmentation of chip pads are important tasks to ensure chip alignment and position correction. To address the challenges of small target chip pad detection, segmentation accuracy and model lightweight, this paper proposes a lightweight chip pad instance segmentation algorithm based on an improved YOLOv8-seg, named YOLOv8-seg-CP (chip pad). Firstly, we integrate the next-generation lightweight StarNet into the original backbone network to enhance fine feature capture capabilities while reducing the number of parameters. Then, we construct the C2f-Star module in the neck network, which enhances the feature extraction performance for small chip pad targets. This maintains accuracy, reduces computational load, and improves detection and segmentation speed. On this basis, we introduce a lightweight shared convolution segmentation head (LSCSH), significantly reducing both parameter count and computational load while enhancing segmentation performance. Additionally, we propose a CGCAFusion convolutional attention fusion module. This module uses a content-guided convolutional attention fusion mechanism to dynamically adjust attention weights based on the content of input features, capturing both global and local feature information and enhancing multimodal feature fusion. Experiments on the chip pad dataset demonstrate that our algorithm achieves a detection and segmentation accuracy of 89.8%. The model size, parameters, and FLOPs are 3.7M, 1.7M, and 8.2G respectively, representing reductions of 45.6%, 50%, and 31.7% compared to the baseline model. The FPS is 1399.3, an improvement of 25.8% over the baseline model. The inference time is 0.72ms, which is 0.18ms less than the baseline model. Extensive experimental results on COCO, carparts-seg, and crack-seg datasets further show that the improved YOLOv8n-seg model outperforms many existing advanced methods in terms of performance. This approach holds significant industrial application value for fully automated chip testing and sorting integrated production.

A Lightweight Cotton Field Weed Detection Model Enhanced with EfficientNet and Attention Mechanisms

Cotton is a crucial crop in the global textile industry, with major production regions including China, India, and the United States. While smart agricultural mechanization technologies, such as automated irrigation and precision pesticide systems, have improved crop management, weeds remain a significant challenge. These weeds not only compete with cotton for nutrients but can also serve as hosts for diseases, affecting both cotton yield and quality. Existing weed detection models perform poorly in the complex environment of cotton fields, where the visual features of weeds and crops are similar and often overlap, resulting in low detection accuracy. Furthermore, real-time deployment on edge devices is difficult. To address these issues, this study proposes an improved lightweight weed detection model, YOLO-WL, based on the YOLOv8 architecture. The model leverages EfficientNet to reconstruct the backbone, reducing model complexity and enhancing detection speed. To compensate for any performance loss due to backbone simplification, CA (cross-attention) is introduced into the backbone, improving feature sensitivity. Finally, AFPN (Adaptive Feature Pyramid Network) and EMA (efficient multi-scale attention) mechanisms are integrated into the neck to further strengthen feature extraction and improve weed detection accuracy. At the same time, the model maintains a lightweight design suitable for deployment on edge devices. Experiments on the CottonWeedDet12 dataset show that the YOLO-WL model achieved an mAP of 92.30%, reduced the detection time per image by 75% to 1.9 ms, and decreased the number of parameters by 30.3%. After TensorRT optimization, the video inference time was reduced from 23.134 ms to 2.443 ms per frame, enabling real-time detection in practical agricultural environments.

Combination of edge enhancement and cold diffusion model for low dose CT image denoising.

Since the quality of low dose CT (LDCT) images is often severely affected by noise and artifacts, it is very important to maintain high quality CT images while effectively reducing the radiation dose. In recent years, the representation of diffusion models to produce high quality images and stable trainability has attracted wide attention. With the extension of the cold diffusion model to the classical diffusion model, its application has greater flexibility. Inspired by the cold diffusion model, we proposes a low dose CT image denoising method, called CECDM, based on the combination of edge enhancement and cold diffusion model. The LDCT image is taken as the end point (forward) of the diffusion process and the starting point (reverse) of the sampling process. Improved sobel operator and Convolution Block Attention Module are added to the network, and compound loss function is adopted. The experimental results show that CECDM can effectively remove noise and artifacts from LDCT images while the inference time of a single image is reduced to 0.41 s. Compared with the existing LDCT image post-processing methods, CECDM has a significant improvement in all indexes.

Prediction of laser beam spatial profiles in a high-energy laser facility by use of deep learning

We adapt the significant advances achieved recently in the field of generative artificial intelligence/machine-learning to laser performance modeling in multipass, high-energy laser systems with application to high-shot-rate facilities relevant to inertial fusion energy. Advantages of neural-network architectures include rapid prediction capability, data-driven processing, and the possibility to implement such architectures within future low-latency, low-power consumption photonic networks. Four models were investigated that differed in their generator loss functions and utilized the U-Net encoder/decoder architecture with either a reconstruction loss alone or combined with an adversarial network loss. We achieved inference times of 1.3 ms for a 256 × 256 pixel near-field beam with errors in predicted energy of the order of 1% over most of the energy range. It is shown that prediction errors are significantly reduced by ensemble averaging the models with different weight initializations. These results suggest that including the temporal dimension in such models may provide accurate, real-time spatiotemporal predictions of laser performance in high-shot-rate laser systems.

Edge Computing for AI-Based Brain MRI Applications: A Critical Evaluation of Real-Time Classification and Segmentation.

Medical imaging plays a pivotal role in diagnostic medicine with technologies like Magnetic Resonance Imagining (MRI), Computed Tomography (CT), Positron Emission Tomography (PET), and ultrasound scans being widely used to assist radiologists and medical experts in reaching concrete diagnosis. Given the recent massive uplift in the storage and processing capabilities of computers, and the publicly available big data, Artificial Intelligence (AI) has also started contributing to improving diagnostic radiology. Edge computing devices and handheld gadgets can serve as useful tools to process medical data in remote areas with limited network and computational resources. In this research, the capabilities of multiple platforms are evaluated for the real-time deployment of diagnostic tools. MRI classification and segmentation applications developed in previous studies are used for testing the performance using different hardware and software configurations. Cost-benefit analysis is carried out using a workstation with a NVIDIA Graphics Processing Unit (GPU), Jetson Xavier NX, Raspberry Pi 4B, and Android phone, using MATLAB, Python, and Android Studio. The mean computational times for the classification app on the PC, Jetson Xavier NX, and Raspberry Pi are 1.2074, 3.7627, and 3.4747 s, respectively. On the low-cost Android phone, this time is observed to be 0.1068 s using the Dynamic Range Quantized TFLite version of the baseline model, with slight degradation in accuracy. For the segmentation app, the times are 1.8241, 5.2641, 6.2162, and 3.2023 s, respectively, when using JPEG inputs. The Jetson Xavier NX and Android phone stand out as the best platforms due to their compact size, fast inference times, and affordability.

LightCardiacNet: light-weight deep ensemble network with attention mechanism for cardiac sound classification

Cardiovascular diseases (CVDs) account for about 32% of global deaths. While digital stethoscopes can record heart sounds, expert analysis is often lacking. To address this, we propose LightCardiacNet, an interpretable, lightweight ensemble neural network using Bi-Directional Gated Recurrent Units (Bi-GRU). It is trained on the PASCAL Heart Challenge and CirCor DigiScope datasets. Static network pruning enhances model sparsity for real-time deployment. We employ various data augmentation techniques to improve resilience to background noise. An ensemble of the two networks is constructed by employing a weighted average approach that combines the two light-weight attention Bi-GRU networks trained on different datasets, which outperforms several state-of-the-art networks achieving an accuracy of 99.8%, specificity of 99.6%, sensitivity of 95.2%, ROC-AUC of 0.974 and inference time of 17 ms on the PASCAL dataset, accuracy of 98.5%, specificity of 95.1%, sensitivity of 90.9%, ROC-AUC of 0.961 and inference time of 18 ms on the CirCor dataset, and an accuracy of 96.21%, sensitivity of 92.78%, specificity of 93.16%, ROC-AUC of 0.913 and inference time of 17.5 ms on real-world data. We adopt the SHAP algorithm to incorporate model interpretability and provide insights to make it clinically explainable and useful to healthcare professionals.

FireNet: A Lightweight and Efficient Multi-Scenario Fire Object Detector

Fire and smoke detection technologies face challenges in complex and dynamic environments. Traditional detectors are vulnerable to background noise, lighting changes, and similar objects (e.g., clouds, steam, dust), leading to high false alarm rates. Additionally, they struggle with detecting small objects, limiting their effectiveness in early fire warnings and rapid responses. As real-time monitoring demands grow, traditional methods often fall short in smart city and drone applications. To address these issues, we propose FireNet, integrating a simplified Vision Transformer (RepViT) to enhance global feature learning while reducing computational overhead. Dynamic snake convolution (DSConv) captures fine boundary details of flames and smoke, especially in complex curved edges. A lightweight decoupled detection head optimizes classification and localization, ideal for high inter-class similarity and small targets. FireNet outperforms YOLOv8 on the Fire Scene dataset (FSD) with a mAP@0.5 of 80.2%, recall of 78.4%, and precision of 82.6%, with an inference time of 26.7 ms. It also excels on the FSD dataset, addressing current fire detection challenges.

Closing the gap between SGP4 and high-precision propagation via differentiable programming

The simplified general perturbations 4 (SGP4) orbital propagation model is one of the most widely used methods for rapidly and reliably predicting the positions and velocities of objects orbiting Earth. Over time, SGP models have undergone refinement to enhance their efficiency and accuracy. Nevertheless, they still do not match the precision offered by high-precision numerical propagators, which can predict the positions and velocities of space objects in low-Earth orbit with significantly smaller errors.In this study, we introduce a novel differentiable version of SGP4, named ∂SGP4. By porting the source code of SGP4 into a differentiable program based on PyTorch, we unlock a whole new class of techniques enabled by differentiable orbit propagation, including spacecraft orbit determination, state conversion, covariance similarity transformation, state transition matrix computation, and covariance propagation. Besides differentiability, our ∂SGP4 supports parallel propagation of a batch of two-line elements (TLEs) in a single execution and it can harness modern hardware accelerators like GPUs or XLA devices (e.g. TPUs) thanks to running on the PyTorch backend.Furthermore, the design of ∂SGP4 makes it possible to use it as a differentiable component in larger machine learning (ML) pipelines, where the propagator can be an element of a larger neural network that is trained or fine-tuned with data. Consequently, we propose a novel orbital propagation paradigm, ML-∂SGP4. In this paradigm, the orbital propagator is enhanced with neural networks attached to its input and output. Through gradient-based optimization, the parameters of this combined model can be iteratively refined to achieve precision surpassing that of SGP4. Fundamentally, the neural networks function as identity operators when the propagator adheres to its default behavior as defined by SGP4. However, owing to the differentiability ingrained within ∂SGP4, the model can be fine-tuned with ephemeris data to learn corrections to both inputs and outputs of SGP4. This augmentation enhances precision while maintaining the same computational speed of ∂SGP4 at inference time. This paradigm empowers satellite operators and researchers, equipping them with the ability to train the model using their specific ephemeris or high-precision numerical propagation data.

Comparative performance of YOLOv8, YOLOv9, YOLOv10, YOLOv11 and Faster R-CNN models for detection of multiple weed species

Weeds pose a serious production challenge in various agronomic crops by reducing their grain yields. Increasing cases of herbicide-resistant (HR) weed populations further exacerbate the problem. Future weed control tactics require the integration of non-chemical and reduced chemical-based strategies that can target site- and specie-specific weed management (SSSWM). Advanced machine learning technology has the potential to localize and detect weed seedlings to implement SSSWM. However, due to large biological variability among various weed species and environmental conditions where they grow, accurate and precise weed detection remains challenging. The main objectives of this research were to (1) develop an annotated image database of cocklebur (Xanthium strumarium L.), dandelion (Taraxacum officinale), common waterhemp (Amaranthus tuberculatus), Palmer amaranth (Amaranthus palmeri) and common lambsquarters (Chenopodium album L.), and (2) investigate the comparative performance (speed and accuracy) of YOLOv8, YOLOv9, YOLOv10, YOLOv11 and Faster R-CNN algorithms in detecting those weed species. A weed dataset with bounding box annotations for each weed species was created, consisting of images collected under variable field conditions which were preprocessed and augmented to create a dataset of 2348 color images. The YOLOv8, YOLOv9, YOLOv10, YOLOv11 and Faster R-CNN were trained using the annotated weed image database to detect each weed species. Results indicated that YOLOv11 was the fastest model with inference time of 13.5 milliseconds (ms) followed by YOLOv8 and YOLOv10 with inference time 23 and 19.3 milliseconds (ms), respectively. The YOLOv9 had the highest accuracy in detecting different weed species with an overall mean average precision (mAP@0.5) of 0.935. In contrast, the Detectron2 with Fast R-CNN configuration provided mAP@0.5 of 0.821 with an inference time of 63.8 ms. These results suggest that the YOLO series algorithms have the potential for real-time deployment for weed species detection more accurately and faster than Faster R-CNN in agricultural fields.

Advanced detection of foreign objects in fresh-cut vegetables using YOLOv5

The presence of foreign objects in fresh-cut vegetables has become a significant concern for the industry in recent years. Ensuring the safety of consumers and suppliers necessitates comprehensive precautionary measures. This study introduces a novel approach for detecting foreign objects in fresh-cut vegetables using YOLOv5. The results indicate that YOLOv5s excels in identifying foreign objects with a high detection accuracy of 98.30%, a rapid inference time of 2.6 ms, and a compact model size of 13.3 MB. The model effectively identified foreign objects such as transparent and colored plastic, paper, wood, stone, insects, glass, and metal. Moreover, YOLOv5s accurately detected foreign objects with color similarities to green onions, which is particularly challenging due to their wide color variation. Hence, the foreign objects dataset was tested and generated 98.63% and 98.67% for cabbage and green onion, respectively. Additionally, YOLOv5s successfully detected small foreign objects (2–3 mm) in two types of fresh-cut vegetables. Despite its excellent performance, the YOLOv5s model struggles to identify foreign objects that overlap with vegetable samples. To address this issue, installing an automatic conveyor unit could facilitate continuous sample movement, while a feeder unit could reduce the possibility of overlap. This research demonstrates the feasibility of implementing the YOLOv5s, as a non-destructive technique for detecting foreign objects in fresh-cut vegetables. The findings contribute to the development of an accurate, fast, and efficient real-time inspection system, with potential applications in the fresh-cut vegetable industry to enhance product quality and safety.

PFEI-Net: A profound feature exploration and interaction network for ceramic substrate surface defect detection

Ceramic substrates serve as the foundational material for numerous electronic devices, and their surface quality directly affects performance and longevity. Therefore, surface defect detection on ceramic substrates is an indispensable task during the manufacturing process. Ceramic substrate surface defects exhibit low contrast, along with intraclass difference and interclass similarity, posing substantial challenges for automated visual inspection. Existing deep learning-based methods utilize carefully crafted backbone networks and feature pyramid structures that perceive multilevel information to deal with the above challenges. However, these methods still suffer from insufficient extraction of discriminative features and low effectiveness in feature fusion, thereby hindering the detection performance. To address these issues, we propose the profound feature exploration and interaction network (PFEI-Net). Initially, we propose the discriminative feature mining backbone (DFM-backbone) to progressively explore the effective discriminative features. In the DFM-backbone, the global information grated perception module is devised to construct long-range feature dependencies and aggregate rich features, whereas the contextual semantic flexible extraction module is designed to meticulously perceive contextual semantic information in a targeted manner, enhancing the comprehensive representation of the discriminative features. Then, we propose the multipath semantic interaction guided-feature pyramid network (MSIG-FPN) to facilitate the fusion and interaction of the valuable information. In the MSIG-FPN, the hierarchical refocus–reaggregation module is constructed to refocus on task-relevant features and provide reliable guidance for the network, whereas the cross-stage semantic deep fusion module is designed to integrate multi-level contextual information deeply across stages and is employed to construct the semantic complementary paths, thereby preventing the loss of crucial features. Finally, built upon the backbone of DFM-backbone and MSIG-FPN, we establish the PFEI-Net, achieving accurate detection of surface defects on ceramic substrates. Experimental results highlight that the PFEI-Net achieves excellent performance with a remarkable 89.3 % mean average precision (mAP) and an inference time of 26.1 ms, which offers a promising and viable solution for automated surface defect detection of ceramic substrate.

DeployFusion: A Deployable Monocular 3D Object Detection with Multi-Sensor Information Fusion in BEV for Edge Devices.

To address the challenges of suboptimal remote detection and significant computational burden in existing multi-sensor information fusion 3D object detection methods, a novel approach based on Bird's-Eye View (BEV) is proposed. This method utilizes an enhanced lightweight EdgeNeXt feature extraction network, incorporating residual branches to address network degradation caused by the excessive depth of STDA encoding blocks. Meantime, deformable convolution is used to expand the receptive field and reduce computational complexity. The feature fusion module constructs a two-stage fusion network to optimize the fusion and alignment of multi-sensor features. This network aligns image features to supplement environmental information with point cloud features, thereby obtaining the final BEV features. Additionally, a Transformer decoder that emphasizes global spatial cues is employed to process the BEV feature sequence, enabling precise detection of distant small objects. Experimental results demonstrate that this method surpasses the baseline network, with improvements of 4.5% in the NuScenes detection score and 5.5% in average precision for detection objects. Finally, the model is converted and accelerated using TensorRT tools for deployment on mobile devices, achieving an inference time of 138 ms per frame on the Jetson Orin NX embedded platform, thus enabling real-time 3D object detection.

Adaptive Transformer-Based Multi-Modal Image Fusion for Real-Time Medical Diagnosis and Object Detection

In recent years, medical diagnosis and object detection have been significantly enhanced by the integration of multi-modal image fusion techniques. This study proposes an Adaptive Transformer-Based Multi-Modal Image Fusion (AT-MMIF) framework designed for real-time medical diagnosis and object detection. The framework employs a Transformer architecture to capture both global and local feature correlations across multiple imaging modalities, including MRI, CT, PET, and X-ray, for more accurate diagnostic results and faster object detection in medical imagery. The fusion process incorporates spatial and frequency-domain information to improve the clarity and detail of the output images, enhancing diagnostic accuracy. The adaptive attention mechanism within the Transformer dynamically adjusts to the relevant features of different image types, optimizing fusion in real time. This leads to an improved sensitivity (98.5%) and specificity (96.7%) in medical diagnosis. Additionally, the model significantly reduces false positives and negatives, with an F1 score of 97.2% in object detection tasks. The AT-MMIF framework is further optimized for real-time processing with an average inference time of 120 ms per image and a model size reduction of 35% compared to existing multi-modal fusion models. By leveraging the strengths of Transformer architectures and adaptive learning, the proposed framework offers a highly efficient and scalable solution for real-time medical diagnosis and object detection in various clinical settings, including radiology, oncology, and pathology.