Complex Scenes Research Articles

Recognizing objects in complex scenes: A Recent Systematic Review

This systematic review provides a comprehensive examination of recent advancements in object recognition within complex scenes, focusing on addressing challenges such as clutter, occlusion, and diverse environmental conditions. Leveraging the PRISMA framework, the study meticulously analysed a diverse range of literature from esteemed sources, employing advanced search methods on Scopus and WoS databases to identify and analyse primary research studies (n = 25). The review encompasses three key themes: Theme 1 concentrates on Image Noise Identification and Removal Techniques, while Theme 2 delves into Image Classification and Recognition under Noise. Additionally, Theme 3 explores Innovative Models and Approaches for Noise-Robust Image Analysis. Despite the progress achieved, contemporary recognition systems struggle with real-world complexities such as varied scales, lighting variations, and different viewpoints. The synthesis of findings emphasizes the necessity for innovative strategies that capitalize on contextual cues and harness the potential of deep learning to enhance precision in object recognition within intricate visual environments. The insights gleaned from this synthesis are poised to guide future research directions, informing the development of more resilient algorithms capable of navigating challenges and catalysing advancements in the field of object recognition.

Fluorescence Microscopic Image Enhancement Method Based on Multi-saliency Guided Filtering Fusion

Abstract Aiming at addressing the issues of colour deviation and clarity reduction in digital pathological imaging systems caused by complex environmental interferences such as uneven slice thickness, sample jitter, and background noise, this paper proposes a multi-saliency-guided filtering fusion enhancement method that combines energy saliency and Gaussian saliency. Firstly, a series of Gamma correction and spatial linear saturation adjustments are applied to enhance the contrast and colour saturation of blurred images, resulting in a sequence of pseudo-exposure images with improved contrast and saturation. Then, the pseudo-exposure images are decomposed into base and detail layers, and the fusion rules of energy saliency and Gaussian saliency-guided filtering are applied to each layer separately. Finally, clear microscopic images are obtained through two-scale reconstruction. This method overcomes the limitations of conventional deblurring algorithms that rely on prior estimation of scene depth and complex depth mapping processes. Experimental results on a dataset of cytological bright-field microscopic imaging samples show that the clarity evaluation metrics, including peak signal-to-noise ratio, root mean square error, and structural similarity index, have significantly improved, indicating that the proposed method can effectively improve the colour fidelity of the samples while eliminating noise interference and enhancing the clarity of cellular texture structures.

CSSDet: small object detection via cross-scale feature enhancement on drone-view images

ABSTRACT Object detection on drone-view images is vital for applications like intelligent transportation, abnormal behavior detection, and urban surveillance. However, the diverse perspectives and altitudes from which Unmanned Aerial Vehicle (UAV) capture scenes pose challenges due to multi-scale differences and small object sizes. To address this, we propose CSSDet, emphasizing multi-scale object detection and small object enhancement. CSSDet integrates an additional detection head for small objects and employs a bidirectional weighted feature pyramid network for effective cross-scale feature fusion, enhancing global perceptual capability. Additionally, we introduce coordinate attention and Involution modules to mitigate information loss. Experiments are conducted on two publicly available datasets, VisDrone and UAVDT. Quantitative results from the experiments indicate that the proposed CSSDet achieves optimal performance across a wide range of evaluation metrics, with particular strengths demonstrated in multi-scale object detection and small-size object detection. Qualitative results showcase its remarkable effectiveness in various aspects of object detection, including small object detection, objects at different scales and categories, multi-perspective scenarios, and complex scenes. The substantial experimental findings confirm that CSSDet attains comprehensive object detection performance on drone-view images. The source code is available at https://github.com/Gui-Cheng/CSSDet.

Robotic MAG welding defects and quality assessment with a defect threshold decision model-driven method

The use of machine vision and deep learning methods for online monitoring and evaluation of welding quality during the welding process and timely detection and correction of welding defects is essential for intelligent welding manufacturing. In this paper, with the goal of online monitoring of the welding quality of the robotic arc welding for the excavator boom, we established a multi-source sensing system for robotic arc welding processes. We have developed various feature extraction algorithms for extracting information about welding pools, arc sounds, currents, voltages, and other features. We proposed to extract welding pool image features using Res2-MobileNetV3 to obtain a 12-dimensional welding pool feature, which enabled good inter-class discrimination among different welding defects. We designed a welding defect threshold decision (DD) strategy model using the responsive machine learning and lightweight network Res2-MobileNetV3 method, achieving a welding defect recognition accuracy of 96.59%. Finally, we tested the accuracy and reliability of the welding defect recognition model on an actual welding scene for the robotic welding of the excavator boom. It provided valuable scientific methods and technological approaches for intelligent welding in complex scenes.

A Computationally Efficient Neuronal Model for Collision Detection with Contrast Polarity-Specific Feed-Forward Inhibition

Animals utilize their well-evolved dynamic vision systems to perceive and evade collision threats. Driven by biological research, bio-inspired models based on lobula giant movement detectors (LGMDs) address certain gaps in constructing artificial collision-detecting vision systems with robust selectivity, offering reliable, low-cost, and miniaturized collision sensors across various scenes. Recent progress in neuroscience has revealed the energetic advantages of dendritic arrangements presynaptic to the LGMDs, which receive contrast polarity-specific signals on separate dendritic fields. Specifically, feed-forward inhibitory inputs arise from parallel ON/OFF pathways interacting with excitation. However, none of the previous research has investigated the evolution of a computational LGMD model with feed-forward inhibition (FFI) separated by opposite polarity. This study fills this vacancy by presenting an optimized neuronal model where FFI is divided into ON/OFF channels, each with distinct synaptic connections. To align with the energy efficiency of biological systems, we introduce an activation function associated with neural computation of FFI and interactions between local excitation and lateral inhibition within ON/OFF channels, ignoring non-active signal processing. This approach significantly improves the time efficiency of the LGMD model, focusing only on substantial luminance changes in image streams. The proposed neuronal model not only accelerates visual processing in relatively stationary scenes but also maintains robust selectivity to ON/OFF-contrast looming stimuli. Additionally, it can suppress translational motion to a moderate extent. Comparative testing with state-of-the-art based on ON/OFF channels was conducted systematically using a range of visual stimuli, including indoor structured and complex outdoor scenes. The results demonstrated significant time savings in silico while retaining original collision selectivity. Furthermore, the optimized model was implemented in the embedded vision system of a micro-mobile robot, achieving the highest success ratio of collision avoidance at 97.51% while nearly halving the processing time compared with previous models. This highlights a robust and parsimonious collision-sensing mode that effectively addresses real-world challenges.

Remote Sensing LiDAR and Hyperspectral Classification with Multi-Scale Graph Encoder–Decoder Network

The rapid development of sensor technology has made multi-modal remote sensing data valuable for land cover classification due to its diverse and complementary information. Many feature extraction methods for multi-modal data, combining light detection and ranging (LiDAR) and hyperspectral imaging (HSI), have recognized the importance of incorporating multiple spatial scales. However, effectively capturing both long-range global correlations and short-range local features simultaneously on different scales remains a challenge, particularly in large-scale, complex ground scenes. To address this limitation, we propose a multi-scale graph encoder–decoder network (MGEN) for multi-modal data classification. The MGEN adopts a graph model that maintains global sample correlations to fuse multi-scale features, enabling simultaneous extraction of local and global information. The graph encoder maps multi-modal data from different scales to the graph space and completes feature extraction in the graph space. The graph decoder maps the features of multiple scales back to the original data space and completes multi-scale feature fusion and classification. Experimental results on three HSI-LiDAR datasets demonstrate that the proposed MGEN achieves considerable classification accuracies and outperforms state-of-the-art methods.

Spatiotemporal Information, Near-Field Perception, and Service for Tourists by Distributed Camera and BeiDou Positioning System in Mountainous Scenic Areas

The collaborative use of camera near-field sensors for monitoring the number and status of tourists is a crucial aspect of smart scenic spot management. This paper proposes a near-field perception technical system that achieves dynamic and accurate detection of tourist targets in mountainous scenic areas, addressing the challenges of real-time passive perception and safety management of tourists. The technical framework involves the following steps: Firstly, real-time video stream signals are collected from multiple cameras to create a distributed perception network. Then, the YOLOX network model is enhanced with the CBAM module and ASFF method to improve the dynamic recognition of preliminary tourist targets in complex scenes. Additionally, the BYTE target dynamic tracking algorithm is employed to address the issue of target occlusion in mountainous scenic areas, thereby enhancing the accuracy of model detection. Finally, the video target monocular spatial positioning algorithm is utilized to determine the actual geographic location of tourists based on the image coordinates. The algorithm was deployed in the Tianmeng Scenic Area of Yimeng Mountain in Shandong Province, and the results demonstrate that this technical system effectively assists in accurately perceiving and spatially positioning tourists in mountainous scenic spots. The system demonstrates an overall accuracy in tourist perception of over 90%, with spatial positioning errors less than 1.0 m and a root mean square error (RMSE) of less than 1.14. This provides auxiliary technical support and effective data support for passive real-time dynamic precise perception and safety management of regional tourist targets in mountainous scenic areas with no/weak satellite navigation signals.

Robust and accurate visual geo-localization using prior map constructed by handheld LiDAR SLAM with camera image and terrestrial LiDAR point cloud

ABSTRACT Visual geo-localization using prior map of known environments has extensive applications in the fields such as self-driving, augmented reality and navigation. Currently, such prior maps are usually constructed by visual SLAM or SFM. However, the recent advances in LiDAR SLAM technology with RGB camera demonstrate great potential in efficient and accurate prior map construction and visual geo-localization. In this research, we developed a novel prior map construction approach which seamlessly integrates a handheld LiDAR SLAM system, a global LiDAR localization algorithm and a sparse set of terrestrial LiDAR scans as accurate control within a factor graph optimization framework. The developed LiDAR SLAM system augmented by terrestrial LiDAR scans achieves an improved mapping accuracy of 8 cm under GNSS-denied conditions in the areas over 50,000 m2 while the state-of-art FAST-LIO2 alone fails and produces a mapping error over 30 m. The results show that a competitive accuracy of visual geo-localization on a mobile phone using the constructed prior map is achieved (i.e. 60 and 76 cm in two tests respectively). Finally, this work also demonstrates a novel LiDAR point cloud fusion method that produces optimized and consistent coarse-to-fine 3D reconstruction in large and complex scenes.

Vehicle Localization Method in Complex SAR Images Based on Feature Reconstruction and Aggregation.

Due to the small size of vehicle targets, complex background environments, and the discrete scattering characteristics of high-resolution synthetic aperture radar (SAR) images, existing deep learning networks face challenges in extracting high-quality vehicle features from SAR images, which impacts vehicle localization accuracy. To address this issue, this paper proposes a vehicle localization method for SAR images based on feature reconstruction and aggregation with rotating boxes. Specifically, our method first employs a backbone network that integrates the space-channel reconfiguration module (SCRM), which contains spatial and channel attention mechanisms specifically designed for SAR images to extract features. The network then connects a progressive cross-fusion mechanism (PCFM) that effectively combines multi-view features from different feature layers, enhancing the information content of feature maps and improving feature representation quality. Finally, these features containing a large receptive field region and enhanced rich contextual information are input into a rotating box vehicle detection head, which effectively reduces false alarms and missed detections. Experiments on a complex scene SAR image vehicle dataset demonstrate that the proposed method significantly improves vehicle localization accuracy. Our method achieves state-of-the-art performance, which demonstrates the superiority and effectiveness of the proposed method.

Hyperspectral Image Classification Algorithm for Forest Analysis Based on a Group-Sensitive Selective Perceptual Transformer

Substantial advancements have been achieved in hyperspectral image (HSI) classification through contemporary deep learning techniques. Nevertheless, the incorporation of an excessive number of irrelevant tokens in large-scale remote sensing data results in inefficient long-range modeling. To overcome this hurdle, this study introduces the Group-Sensitive Selective Perception Transformer (GSAT) framework, which builds upon the Vision Transformer (ViT) to enhance HSI classification outcomes. The innovation of the GSAT architecture is primarily evident in several key aspects. Firstly, the GSAT incorporates a Group-Sensitive Pixel Group Mapping (PGM) module, which organizes pixels into distinct groups. This allows the global self-attention mechanism to function within these groupings, effectively capturing local interdependencies within spectral channels. This grouping tactic not only boosts the model’s spatial awareness but also lessens computational complexity, enhancing overall efficiency. Secondly, the GSAT addresses the detrimental effects of superfluous tokens on model efficacy by introducing the Sensitivity Selection Framework (SSF) module. This module selectively identifies the most pertinent tokens for classification purposes, thereby minimizing distractions from extraneous information and bolstering the model’s representational strength. Furthermore, the SSF refines local representation through multi-scale feature selection, enabling the model to more effectively encapsulate feature data across various scales. Additionally, the GSAT architecture adeptly represents both global and local features of HSI data by merging global self-attention with local feature extraction. This integration strategy not only elevates classification precision but also enhances the model’s versatility in navigating complex scenes, particularly in urban mapping scenarios where it significantly outclasses previous deep learning methods. The advent of the GSAT architecture not only rectifies the inefficiencies of traditional deep learning approaches in processing extensive remote sensing imagery but also markededly enhances the performance of HSI classification tasks through the deployment of group-sensitive and selective perception mechanisms. It presents a novel viewpoint within the domain of hyperspectral image classification and is poised to propel further advancements in the field. Empirical testing on six standard HSI datasets confirms the superior performance of the proposed GSAT method in HSI classification, especially within urban mapping contexts, where it exceeds the capabilities of prior deep learning techniques. In essence, the GSAT architecture markedly refines HSI classification by pioneering group-sensitive pixel group mapping and selective perception mechanisms, heralding a significant breakthrough in hyperspectral image processing.

Image Captioning Based on Semantic Scenes.

With the development of artificial intelligence and deep learning technologies, image captioning has become an important research direction at the intersection of computer vision and natural language processing. The purpose of image captioning is to generate corresponding natural language descriptions by understanding the content of images. This technology has broad application prospects in fields such as image retrieval, autonomous driving, and visual question answering. Currently, many researchers have proposed region-based image captioning methods. These methods generate captions by extracting features from different regions of an image. However, they often rely on local features of the image and overlook the understanding of the overall scene, leading to captions that lack coherence and accuracy when dealing with complex scenes. Additionally, image captioning methods are unable to extract complete semantic information from visual data, which may lead to captions with biases and deficiencies. Due to these reasons, existing methods struggle to generate comprehensive and accurate captions. To fill this gap, we propose the Semantic Scenes Encoder (SSE) for image captioning. It first extracts a scene graph from the image and integrates it into the encoding of the image information. Then, it extracts a semantic graph from the captions and preserves semantic information through a learnable attention mechanism, which we refer to as the dictionary. During the generation of captions, it combines the encoded information of the image and the learned semantic information to generate complete and accurate captions. To verify the effectiveness of the SSE, we tested the model on the MSCOCO dataset. The experimental results show that the SSE improves the overall quality of the captions. The improvement in scores across multiple evaluation metrics further demonstrates that the SSE possesses significant advantages when processing identical images.

Railway foreign object tracking and detection with spatial positioning and feature generalization enhancement

The existing deep learning foreign object tracking and detection algorithm is easily affected by complex environments and target occlusion, resulting in problems such as missed detection and low detection accuracy. A railway foreign object tracking and detection algorithm with spatial positioning and feature generalization enhancement is proposed. First, a multi-scale cascade improved GhostNet feature network is proposed to improve the feature extraction capability of infrared targets. Secondly, the spatial positioning and feature generalization enhancement module is designed by using the spatial position positioning and generalized morphological information of foreign objects to enhance the detection accuracy of targets with position movement and tracking trajectory changesin complex scenes. Then, a pyramid prediction network is constructed to obtain the detection anchor frame, type and confidence information of infrared railway foreign objects. Finally, by improving the DeepSORT tracking algorithm with improved category and confidence display, combined with Kalman filtering and Hungarian algorithm, railway foreign object tracking and detection in infrared weak light environment is realized. Experimental results show that the proposed method has a tracking and detection accuracy of railway foreign objects 83.3%, with an average detection rate of 11.3 frames per second. Compared with the comparison method, the proposed method has higher detection accuracy and has better performance in railway foreign object tracking and detection in infrared weak light scenes.

UNeXt: An Efficient Network for the Semantic Segmentation of High-Resolution Remote Sensing Images

The application of deep neural networks for the semantic segmentation of remote sensing images is a significant research area within the field of the intelligent interpretation of remote sensing data. The semantic segmentation of remote sensing images holds great practical value in urban planning, disaster assessment, the estimation of carbon sinks, and other related fields. With the continuous advancement of remote sensing technology, the spatial resolution of remote sensing images is gradually increasing. This increase in resolution brings about challenges such as significant changes in the scale of ground objects, redundant information, and irregular shapes within remote sensing images. Current methods leverage Transformers to capture global long-range dependencies. However, the use of Transformers introduces higher computational complexity and is prone to losing local details. In this paper, we propose UNeXt (UNet+ConvNeXt+Transformer), a real-time semantic segmentation model tailored for high-resolution remote sensing images. To achieve efficient segmentation, UNeXt uses the lightweight ConvNeXt-T as the encoder and a lightweight decoder, Transnext, which combines a Transformer and CNN (Convolutional Neural Networks) to capture global information while avoiding the loss of local details. Furthermore, in order to more effectively utilize spatial and channel information, we propose a SCFB (SC Feature Fuse Block) to reduce computational complexity while enhancing the model’s recognition of complex scenes. A series of ablation experiments and comprehensive comparative experiments demonstrate that our method not only runs faster than state-of-the-art (SOTA) lightweight models but also achieves higher accuracy. Specifically, our proposed UNeXt achieves 85.2% and 82.9% mIoUs on the Vaihingen and Gaofen5 (GID5) datasets, respectively, while maintaining 97 fps for 512 × 512 inputs on a single NVIDIA GTX 4090 GPU, outperforming other SOTA methods.

CGADNet: A Lightweight, Real-Time, and Robust Crosswalk and Guide Arrow Detection Network for Complex Scenes

CGADNet: A Lightweight, Real-Time, and Robust Crosswalk and Guide Arrow Detection Network for Complex Scenes

A generalized model for accurate wheat spike detection and counting in complex scenarios

Wheat is a crucial crop worldwide, and accurate detection and counting of wheat spikes are vital for yield estimation and breeding. However, these tasks are daunting in complex field environments. To tackle this, we introduce RIA-SpikeNet, a model designed to detect and count wheat spikes in such conditions. First, we introduce an Implicit Decoupling Detection Head to incorporate more implicit knowledge, enabling the model to better distinguish visually similar wheat spikes. Second, Asymmetric Loss is employed as the confidence loss function, enhancing the learning weights of positive and hard samples, thus improving performance in complex scenes. Lastly, the backbone network is modified through reparameterization and the use of larger convolutional kernels, expanding the effective receptive field and improving shape information extraction. These enhancements significantly improve the model’s ability to detect and count wheat spikes accurately. RIA-SpikeNet outperforms the state-of-the-art YOLOv8 detection model, achieving a competitive 81.54% mAP and 90.29% R2. The model demonstrates superior performance in challenging scenarios, providing an effective tool for wheat spike yield estimation in field environments and valuable support for wheat production and breeding efforts.

P2FTrack: Multi-object tracking with motion Prior and Feature Posterior

Multiple object tracking (MOT) has emerged as a crucial component of the rapidly developing computer vision. However, existing multi-object tracking methods often overlook the relationship between features and motion, hindering the ability to strike a performance balance between coupled motion and complex scenes. In this work, we propose a novel end-to-end multi-object tracking method that integrates motion and feature information. To achieve this, we introduce a motion prior generator that transforms motion information into attention masks. Additionally, we leverage prior-posterior fusion multi-head attention to combine the motion-derived priors and attention-based posteriors. Our proposed method is extensively evaluated on MOT17 and DanceTrack datasets through comprehensive experiments and ablation studies, demonstrating state-of-the-art performance in the feature-based method with reasonable speed.

A Person Re-Identification Method for Fusing Convolutional Attention and Transformer Architecture

Rapid and accurate localization of target person via person re-identification is one of the important methods in intelligent security systems,but faces great challenges due to complex environmental changes. In order to build a model applicable to various complex scenes, deep learning-based person re-identification models have become a key research direction for researchers. The existing convolutional neural network-based or transformer-based deep learning models are still have certain shortcomings in the extraction of local details and are not able to cope well with realistic complex scenes. To address this problem, a method fusing convolutional neural network and transformer (FCAT) is proposed to enhance the transformer's attention to local detail information. This method mainly improves the transformer's ability to extract local detail features indirectly by embedding convolutional space attention and channel attention respectively to enhance the attention to important regions and important channel features in the image. Comparative ablation experiments on three publicly available person re-identification datasets demonstrate that FCAT achieves comparable results to existing methods on non-occluded datasets, and significantly improves performance on occluded datasets. The proposed model is also has a lightweight effect, and the inference speed is improved without increasing the amount of outer computation and model parameters.

LFIR-YOLO: Lightweight Model for Infrared Vehicle and Pedestrian Detection

The complexity of urban road scenes at night and the inadequacy of visible light imaging in such conditions pose significant challenges. To address the issues of insufficient color information, texture detail, and low spatial resolution in infrared imagery, we propose an enhanced infrared detection model called LFIR-YOLO, which is built upon the YOLOv8 architecture. The primary goal is to improve the accuracy of infrared target detection in nighttime traffic scenarios while meeting practical deployment requirements. First, to address challenges such as limited contrast and occlusion noise in infrared images, the C2f module in the high-level backbone network is augmented with a Dilation-wise Residual (DWR) module, incorporating multi-scale infrared contextual information to enhance feature extraction capabilities. Secondly, at the neck of the network, a Content-guided Attention (CGA) mechanism is applied to fuse features and re-modulate both initial and advanced features, catering to the low signal-to-noise ratio and sparse detail features characteristic of infrared images. Third, a shared convolution strategy is employed in the detection head, replacing the decoupled head strategy and utilizing shared Detail Enhancement Convolution (DEConv) and Group Norm (GN) operations to achieve lightweight yet precise improvements. Finally, loss functions, PIoU v2 and Adaptive Threshold Focal Loss (ATFL), are integrated into the model to better decouple infrared targets from the background and to enhance convergence speed. The experimental results on the FLIR and multispectral datasets show that the proposed LFIR-YOLO model achieves an improvement in detection accuracy of 4.3% and 2.6%, respectively, compared to the YOLOv8 model. Furthermore, the model demonstrates a reduction in parameters and computational complexity by 15.5% and 34%, respectively, enhancing its suitability for real-time deployment on resource-constrained edge devices.

Dictionary trained attention constrained low rank and sparse autoencoder for hyperspectral anomaly detection

Dictionary representations and deep learning Autoencoder (AE) models have proven effective in hyperspectral anomaly detection. Dictionary representations offer self-explanation but struggle with complex scenarios. Conversely, autoencoders can capture details in complex scenes but lack self-explanation. Complex scenarios often involve extensive spatial information, making its utilization crucial in hyperspectral anomaly detection. To effectively combine the advantages of both methods and address the insufficient use of spatial information, we propose an attention constrained low-rank and sparse autoencoder for hyperspectral anomaly detection. This model includes two encoders: an attention constrained low-rank autoencoder (AClrAE) trained with a background dictionary and incorporating a Global Self-Attention Module (GAM) to focus on global spatial information, resulting in improved background reconstruction; and an attention constrained sparse autoencoder (ACsAE) trained with an anomaly dictionary and incorporating a Local Self-Attention Module (LAM) to focus on local spatial information, resulting in enhanced anomaly reconstruction. Finally, to merge the detection results from both encoders, a nonlinear fusion scheme is employed. Experiments on multiple real and synthetic datasets demonstrate the effectiveness and feasibility of the proposed method.

Complex Scene Understanding and Object Detection Algorithm Assisted by Artificial Intelligence

The purpose of this paper is to study the algorithm of object detection and scene understanding in complex scenes assisted by AI(Artificial Intelligence), so as to improve the machine's perception of the surrounding environment. Aiming at occlusion and dense scenes, this paper proposes a series of innovative algorithms by introducing technologies such as multi-scale detection, occlusion processing and real-time optimization. Experimental results show that the proposed algorithm has achieved excellent performance on several public data sets. Compared with the current mainstream object detection algorithm -YOLO, the method proposed in this paper has a significant improvement in accuracy. Especially in multi-scale object detection, occlusion and dense scene processing and real-time optimization, the proposed algorithm shows obvious advantages. The research results provide an effective solution for object detection and scene understanding in complex scenes, and promote the application of AI system in related fields.