IET Computer Vision最新文献

In the current context of the extensive use of deep neural networks, it has been observed that neural network models are vulnerable to adversarial perturbations, which may lead to unexpected results. In this paper, we introduce an Adversarial Purification Model rooted in latent representation compression, aimed at enhancing the robustness of deep learning models. Initially, we employ an encoder-decoder architecture inspired by the U-net to extract features from input samples. Subsequently, these features undergo a process of information compression to remove adversarial perturbations from the latent space. To counteract the model's tendency to overly focus on fine-grained details of input samples, resulting in ineffective adversarial sample purification, an early freezing mechanism is introduced during the encoder training process. We tested our model's ability to purify adversarial samples generated from the CIFAR-10, CIFAR-100, and ImageNet datasets using various methods. These samples were then used to test ResNet, an image recognition classifiers. Our experiments covered different resolutions and attack types to fully assess LRCM's effectiveness against adversarial attacks. We also compared LRCM with other defence strategies, demonstrating its strong defensive capabilities.

Recently, the accuracy of self-supervised deep learning models for indoor depth estimation has approached that of supervised models by improving the supervision in planar regions. However, a common issue with integrating multiple planar priors is the generation of oversmooth depth maps, leading to unrealistic and erroneous depth representations at edges. Despite the fact that edge pixels only cover a small part of the image, they are of high significance for downstream tasks such as visual odometry, where image features, essential for motion computation, are mostly located at edges. To improve erroneous depth predictions at edge regions, we delve into the self-supervised training process, identifying its limitations and using these insights to develop a geometric edge model. Building on this, we introduce a novel algorithm that utilises the smooth depth predictions of existing models and colour image data to accurately identify edge pixels. After finding the edge pixels, our approach generates targeted self-supervision in these zones by interpolating depth values from adjacent planar areas towards the edges. We integrate the proposed algorithms into a novel loss function that encourages neural networks to predict sharper and more accurate depth edges in indoor scenes. To validate our methodology, we incorporated the proposed edge-enhancing loss function into a state-of-the-art self-supervised depth estimation framework. Our results demonstrate a notable improvement in the accuracy of edge depth predictions and a 19% improvement in visual odometry when using our depth model to generate RGB-D input, compared to the baseline model.

The Grévy's zebra, an endangered species native to Kenya and southern Ethiopia, has been the target of sustained conservation efforts in recent years. Accurately monitoring Grévy's zebra populations is essential for ecologists to evaluate ongoing conservation initiatives. Recently, in both 2016 and 2018, a full census of the Grévy's zebra population was enabled by the Great Grévy's Rally (GGR), a citizen science event that combines teams of volunteers to capture data with computer vision algorithms that help experts estimate the number of individuals in the population. A complementary, scalable, cost-effective and long-term Grévy's population monitoring approach involves deploying a network of camera traps, which we have done at the Mpala Research Centre in Laikipia County, Kenya. In both scenarios, a substantial majority of the images of zebras are not usable for individual identification due to ‘in-the-wild’ imaging conditions—occlusions from vegetation or other animals, oblique views, low image quality and animals that appear in the far background and are thus too small to identify. Camera trap images, without an intelligent human photographer to select the framing and focus on the animals of interest, are of even poorer quality, with high rates of occlusion and high spatiotemporal similarity within image bursts. We employ an image filtering pipeline incorporating animal detection, species identification, viewpoint estimation, quality evaluation and temporal subsampling to compensate for these factors and obtain individual crops from camera trap and GGR images of suitable quality for re-ID. We then employ the local clusterings and their alternatives (LCA) algorithm, a hybrid computer vision and graph clustering method for animal re-ID, on the resulting high-quality crops. Our method processed images taken during GGR-16 and GGR-18 in Meru County, Kenya, into 4142 highly comparable annotations, requiring only 120 contrastive same-vs-different-individual decisions from a human reviewer to produce a population estimate of 349 individuals (within 4.6$��\%�$ of the ground truth count in Meru County). Our method also efficiently processed 8.9M unlabelled camera trap images from 70 camera traps at Mpala over 2 years into 685 encounters of 173 unique individuals, requiring only 331 contrastive decisions from a human reviewer.

Smoke often appears before visible flames in the early stages of fire disasters, making accurate pixel-wise detection essential for fire alarms. Although existing segmentation models effectively identify smoke pixels, they generally treat all pixels within a smoke region as having the same prior probability. This assumption of rigidity, common in natural object segmentation, fails to account for the inherent variability within smoke. We argue that pixels within smoke exhibit a probabilistic relationship with both smoke and background, necessitating density estimation to enhance the representation of internal structures within the smoke. To this end, we propose enhancements across the entire network. First, we improve the backbone by adaptively integrating scene information into texture features through separate paths, enabling smoke-tailored feature representation for further exploit. Second, we introduce a texture-aware head with long convolutional kernels to integrate both global and orientation-specific information, enhancing representation for intricate smoke structure. Third, we develop a dual-task decoder for simultaneous density and location recovery, with the frequency-domain alignment in the final stage to preserve internal smoke details. Extensive experiments on synthetic and real smoke datasets demonstrate the effectiveness of our approach. Specifically, comparisons with 17 models show the superiority of our method, with mean IoU improvements of 4.88%, 2.63%, and 3.17% on three test sets. (The code will be available on https://github.com/xia-xx-cv/TANet_smoke).

Vehicle re-identification (Re-ID) facilitates the recognition and distinction of vehicles based on their visual characteristics in images or videos. However, accurately identifying a vehicle poses great challenges due to (i) the pronounced intra-instance variations encountered under varying lighting conditions such as day and night and (ii) the subtle inter-instance differences observed among similar vehicles. To address these challenges, the authors propose Angle Metric learning for Discriminative Features on vehicle Re-ID (termed as AMDF), which aims to maximise the variance between visual features of different classes while minimising the variance within the same class. AMDF comprehensively measures the angle and distance discrepancies between features. First, to mitigate the impact of lighting conditions on intra-class variation, the authors employ CycleGAN to generate images that simulate consistent lighting (either day or night), thereby standardising the conditions for distance measurement. Second, Swin Transformer was integrated to help generate more detailed features. At last, a novel angle metric loss based on cosine distance is proposed, which organically integrates angular metric and 2-norm metric, effectively maximising the decision boundary in angular space. Extensive experimental evaluations on three public datasets including VERI-776, VERI-Wild, and VEHICLEID, indicate that the method achieves state-of-the-art performance. The code of this project is released at https://github.com/ZnCu-0906/AMDF.

Person re-identification (Re-ID) has gained popularity in computer vision, enabling cross-camera pedestrian recognition. Although the development of deep learning has provided a robust technical foundation for person Re-ID research, most existing person Re-ID methods overlook the potential relationships among local person features, failing to adequately address the impact of pedestrian pose variations and local body parts occlusion. Therefore, we propose a transformer-enhanced graph convolutional network (Tran-GCN) model to improve person re-identification performance in monitoring videos. The model comprises four key components: (1) a pose estimation learning branch is utilised to estimate pedestrian pose information and inherent skeletal structure data, extracting pedestrian key point information; (2) a transformer learning branch learns the global dependencies between fine-grained and semantically meaningful local person features; (3) a convolution learning branch uses the basic ResNet architecture to extract the person's fine-grained local features; and (4) a Graph convolutional module (GCM) integrates local feature information, global feature information and body information for more effective person identification after fusion. Quantitative and qualitative analysis experiments conducted on three different datasets (Market-1501, DukeMTMC-ReID and MSMT17) demonstrate that the Tran-GCN model can more accurately capture discriminative person features in monitoring videos, significantly improving identification accuracy.

The bilateral asymmetry of flanks, where the sides of an animal with unique visual markings are independently patterned, complicates tasks such as individual identification. Automatically generating additional information on the visible side of the animal would improve the accuracy of individual identification. In this study, we used transfer learning on popular convolutional neural network (CNN) image classification architectures to train a flank predictor that predicted the visible flank of quadruped mammalian species in images. We automatically derived the data labels from existing datasets initially labelled for animal pose estimation. The developed models were evaluated across various scenarios involving unseen quadruped species in familiar and unfamiliar habitats. As a real-world scenario, we used a dataset of manually labelled Eurasian lynx (Lynx lynx) from camera traps in the Bavarian Forest National Park, Germany, to evaluate the model. The best model on data obtained in the field was trained on a MobileNetV2 architecture. It achieved an accuracy of 91.7% for the unseen/untrained species lynx in a complex unseen/untrained habitat with challenging light conditions. The developed flank predictor was designed to be embedded as a preprocessing step for automated analysis of camera trap datasets to enhance tasks such as individual identification.

Existing click-based interactive image segmentation methods typically initiate object extraction with the first click and iteratively refine the coarse segmentation through subsequent interactions. Unlike box-based methods, click-based approaches mitigate ambiguity when multiple targets are present within a single bounding box, but suffer from a lack of precise location and outline information. Inspired by instance segmentation, the authors propose a Generated-bbox Guided method that provides location and outline information using an automatically generated bounding box, rather than a manually labelled one, minimising the need for extensive user interaction. Building on the success of vision transformers, the authors adopt them as the network architecture to enhance model's performance. A click-based interactive image segmentation network named the Generated-bbox Guided Coarse-to-Fine Network (GCFN) was proposed. GCFN is a two-stage cascade network comprising two sub-networks: Coarsenet and Finenet. A transformer-based Box Detector was introduced to generate an initial bounding box from a inside click, that can provide location and outline information. Additionally, two feature enhancement modules guided by foreground and background information: the Foreground-Background Feature Enhancement Module (FFEM) and the Pixel Enhancement Module (PEM) were designed. The authors evaluate the GCFN method on five popular benchmark datasets and demonstrate the generalisation capability on three medical image datasets.

Active domain adaptation (active DA) provides an effective solution by selectively labelling a limited number of target samples to significantly enhance adaptation performance. However, existing active DA methods often struggle in real-world scenarios where, due to data privacy concerns, only a pre-trained source model is available, rather than the source samples. To address this issue, we propose a novel method called the structure-based uncertainty estimation model (SUEM) for source-free active domain adaptation (SFADA). To be specific, we introduce an innovative active sample selection strategy that combines both uncertainty and diversity sampling to identify the most informative samples. We assess the uncertainty in target samples using structure-wise probabilities and implement a diversity selection method to minimise redundancy. For the selected samples, we not only apply standard-supervised loss but also conduct interpolation consistency training to further explore the structural information of the target domain. Extensive experiments across four widely used datasets demonstrate that our method is comparable to or outperforms current UDA and active DA methods.

Skeleton-based action recognition using Graph Convolutional Networks (GCNs) has achieved remarkable performance, but recognising ambiguous actions, such as ‘waving’ and ‘saluting’, remains a significant challenge. Existing methods typically rely on a serial combination of GCNs and Temporal Convolutional Networks (TCNs), where spatial and temporal features are extracted independently, leading to an unbalanced spatial-temporal information, which hinders accurate action recognition. Moreover, existing methods for ambiguous actions often overemphasise local details, resulting in the loss of crucial global context, which further complicates the task of differentiating ambiguous actions. To address these challenges, the authors propose a lightweight plug-and-play module called Synchronised and Fine-grained Head (SF-Head), inserted between GCN and TCN layers. SF-Head first conducts Synchronised Spatial-Temporal Extraction (SSTE) with a Feature Redundancy Loss (F-RL), ensuring a balanced interaction between the two types of features. It then performs Adaptive Cross-dimensional Feature Aggregation (AC-FA), with a Feature Consistency Loss (F-CL), which aligns the aggregated feature with their original spatial-temporal feature. This aggregation step effectively combines both global context and local details, enhancing the model's ability to classify ambiguous actions. Experimental results on NTU RGB + D 60, NTU RGB + D 120, NW-UCLA and PKU-MMD I datasets demonstrate significant improvements in distinguishing ambiguous actions. Our code will be made available at https://github.com/HaoHuang2003/SFHead.