Search

Multi-scale feature fusion has been widely used in handcrafted descriptors, but has not been fully explored in deep learning-based descriptor extraction. Simple concatenation of descriptors of different scales has not been successful in significantly improving performance for computer vision tasks. In this paper, we propose a novel convolutional neural network, based on center-surround adaptive multi-scale feature fusion. Our approach enables the network to focus on different center-surround scales, resulting in improved performance. We also introduce a novel regularization technique that uses second-order similarity to constrain the learning of local descriptors, based on the symmetric property of the similarity matrix. The proposed method outperforms single-scale or simple-concatenation descriptors on two datasets and achieves state-of-the-art results on the Brown dataset. Furthermore, our method demonstrates excellent generalization ability on the HPatches dataset. Our code is released on GitHub: https://github.com/Leung-GD/AFSRNet/tree/main. [ABSTRACT FROM AUTHOR]

Zero-shot learning aims to learn a visual classifier for a category which has no training samples leveraging its semantic information and its relationship to other categories. It is common, yet vital, in practical visual scenarios, and particularly prominent in the uncharted ocean field. Phytoplankton plays an important part in the marine ecological environment. It is common to encounter the zero-shot recognition problem during the in situ observation. Therefore, we propose a dual autoencoder model, which contains two similar encoder–decoder structures, to tackle the zero-shot recognition problem. The first one is used for the projection from the visual feature space to a latent space, then to the semantic space. Inversely, the second one projects from the semantic space to another latent space, then back to the visual feature space. This structure guarantees the projection from the visual feature space to the semantic space to be more effective, through the stable mutual mapping. Experimental results on four benchmarks demonstrate that the proposed dual autoencoder model achieves competitive performance compared with six recent state-of-the-art methods. Furthermore, we apply our algorithm to phytoplankton classification. We manually annotated phytoplankton attributes to develop a practical dataset for this real and special domain application, i.e., Zero-shot learning dataset for PHYtoplankton (ZeroPHY). Experiment results show that our method achieves the best performance on this real-world application. [ABSTRACT FROM AUTHOR]

Visual perception modelling of saliency detection has received widespread concerns recently from both the cybernetics and computational intelligence domains. In particular, those distinct background and foreground-oriented models are capable of engendering competitive results. The implicitly vital issue of the above computing approaches is how to reliably choose seeds of the foreground and background cues for kicking off the subsequent saliency-detection procedure. To address this barrier, this paper explores the local geometry and statistical attribute of the detected orientational blocks via an improved discrete wavelet frame transform algorithm to estimate the center position of individual salient object in the original input. Specially, the calculated centroid can be regarded as the prominent focus of visual perception in the initial image, which is beneficial to choose the credible seed during the computing of the superpixel-based foreground and background cues. Then, both sides of the complementary and visually oriented cues are integrated concurrently into a dependable and robust saliency map with reliability. Substantial experimental evaluations in term of freely open-access databases testify the effectiveness of the designed framework, and have prove that the designed algorithm is accurate and outperforms the other distinct representative saliency detection algorithms. [ABSTRACT FROM AUTHOR]

Photometric stereo aims to recover the surface normals of a 3D object from various shading cues, establishing the relationship between two-dimensional images and the object geometry. Traditional methods usually adopt simplified reflectance models to approximate the non-Lambertian surface properties, while recently, photometric stereo based on deep learning has been widely used to deal with non-Lambertian surfaces. However, previous studies are limited in dealing with high-frequency surface regions, i.e., regions with rapid shape variations, such as crinkles, edges, etc., resulted in blurry reconstructions. To alleviate this problem, we present a normalized attention-weighted photometric stereo network, namely NormAttention-PSN, to improve surface orientation prediction, especially for those complicated structures. In order to address these challenges, in this paper, we (1) present an attention-weighted loss to produce better surface reconstructions, which applies a higher weight to the detail-preserving gradient loss in high-frequency areas, (2) adopt a double-gate normalization method for non-Lambertian surfaces, to explicitly distinguish whether the high-frequency representation is stimulated by surface structure or spatially varying reflectance, and (3) adopt a parallel high-resolution structure to generate deep features that can maintain the high-resolution details of surface normals. Extensive experiments on public benchmark data sets show that the proposed NormAttention-PSN significantly outperforms traditional calibrated photometric stereo algorithms and state-of-the-art deep learning-based methods. [ABSTRACT FROM AUTHOR]

At present, poor background suppression is one major problem for visual saliency detection. Although many mainstream saliency detection models can effectively locate salient objects, objects in complicated backgrounds in some natural images are often mistaken for salient objects. Therefore, this paper proposed to set up a center prior-based encoder-decoder network to improve background suppression results. A traditional center prior-based method and U-Net model were combined efficiently. First, multi-scale group convolution was used to replace general convolution, which can highlight the semantic information characteristics, and high-level characteristics at the bottom of U-Net were integrated and optimized on the basis of the consideration of center prior. Then, refinements were delivered throughout the whole network by upgrading the network structure, so as to ensure the optimized features can be made full use of. Since the changes to the U-Net architecture somewhat affected the stability of the network. Therefore, branch network modules were adopted and adaptive parameters defined to coordinate the relationships between branch networks to keep the network structure well balanced. The method has been tested with four widely used databases and proven effective by comparing its results with those of another seven popular methods. [ABSTRACT FROM AUTHOR]

Fine-grained recognition is a challenging task due to small intra-category variances. Most of the top-performing fine-grained recognition methods leverage parts of objects for better performance. Therefore, part annotations which are extremely computationally expensive are required. In this paper, we propose a novel cascaded deep CNN detection framework for fine-grained recognition which is trained to detect a whole object without considering parts. Nevertheless, most of the current top-performing detection networks use N + 1 class (N object categories plus background) softmax loss. The background category with much more training samples dominates the feature learning progress where the features are not suitable for object categorisation with fewer samples. To address this issue, we here introduce two strategies: 1) We leverage a cascaded structure to eliminate the background. 2) We introduce a novel one-vs-rest loss function to capture more minute variances from different subordinate categories. Experiments show that our proposed recognition framework achieves comparable performance against the state-of-the-art, part-free, fine-grained recognition methods on the CUB-200-2011 Bird dataset. Meanwhile, our method outperforms most of the existing part annotation based methods and does not need part annotations at the training stage whilst being free from any annotations at the test stage. [ABSTRACT FROM AUTHOR]

Recently, with the development of various camera sensors and internet network, the volume of digital images is becoming big. Content-based image retrieval (CBIR), especially in network big data analysis, has attracted wide attention. CBIR systems normally search the most similar images to the given query example among a wide range of candidate images. However, human psychology suggests that users concern more about regions of their interest and merely want to retrieve images containing relevant regions, while ignoring irrelevant image areas (such as the texture regions or background). Previous CBIR system on user-interested image retrieval generally requires complicated segmentation of the region from the background. In this paper, we propose a novel hierarchical-local-feature extraction scheme for CBIR, whereas complex image segmentation is avoided. In our CBIR system, a perception-based directional patch extraction method and an improved salient patch detection algorithm are proposed for local features extraction. Then, color moments and Gabor texture features are employed to index the salient regions. Extensive experiments have been performed to evaluate the performance of the proposed scheme, and experimental results show that the developed CBIR system produces plausible retrieval results. [ABSTRACT FROM AUTHOR]

In this paper, a novel and efficient framework by exploiting Quaternionic Distance Based Weber Local Descriptor (QDWLD) and object cues is proposed for image saliency detection. In contrast to the existing saliency detection models, the advantage of the proposed approach is that it can combine quaternion number system and object cues simultaneously, which is independent of image contents and scenes. Firstly, QDWLD, which was initially designed for detecting outliers in color images, is used to represent the directional cues in an image. Meanwhile, two low-level priors, namely the Convex-Hull-Based center and color contrast cue of the image, are utilized and fused as an object-level cue. Finally, by combining QDWLD with object cues, a reliable saliency map of the image can be computed and estimated. Experimental results, based on two widely used and openly available database, show that the proposed method is able to produce reliable and promising salient maps/estimations, compared to other state-of-the-art saliency-detection models. [ABSTRACT FROM AUTHOR]

Image classification usually requires complicated segmentation to separate foreground objects from the background scene. However, the statistical content of a background scene can actually provide very useful information for classification. In this paper, we propose a new hybrid pyramid kernel which incorporates local features extracted from both dense regular grids and interest points for image classification, without requiring segmentation. Features extracted from dense regular grids can better capture information about the background scene, while interest points detected at corners and edges can better capture information about the salient objects. In our algorithm, these two local features are combined in both the spatial and the feature-space domains, and are organized into pyramid representations. In order to obtain better classification accuracy, we fine-tune the parameters involved in the similarity measure, and we determine discriminative regions by means of relevance feedback. From the experimental results, we observe that our algorithm can achieve a 6.37 % increase in performance as compared to other pyramid-representation-based methods. To evaluate the applicability of the proposed hybrid kernel to large-scale databases, we have performed a cross-dataset experiment and investigated the effect of foreground/background features on each of the kernels. In particular, the proposed hybrid kernel has been proven to satisfy Mercer's condition and is efficient in measuring the similarity between image features. For instance, the computational complexity of the proposed hybrid kernel is proportional to the number of features. [ABSTRACT FROM AUTHOR]

An image magnification method with a Gradient Vector Flow (GVF) constraint-based anisotropic diffusion model is proposed in this letter. A Low-Resolution (LR) image is first magnified using bilinear interpolation, and then an iterative image restoration method, with the use of an anisotropic diffusion model and a Gaussian moving-average constraint, is applied to the magnified image. The estimated GVF of a High-Resolution (HR) image can be used to remove the jagged effect and to preserve the textural structure in the image. Meanwhile, the use of the Gaussian moving-average LR model can provide a data fidelity constraint, which renders a magnified image closer to the ideal HR version. Experimental results show that the proposed method can improve the quality of magnified images in terms of both objective and subjective criteria. [ABSTRACT FROM AUTHOR]

Underwater image enhancement presents a significant challenge due to the complex and diverse underwater environments that result in severe degradation phenomena such as light absorption, scattering, and color distortion. More importantly, obtaining paired training data for these scenarios is a challenging task, which further hinders the generalization performance of enhancement models. To address these issues, we propose a novel approach, the Hybrid Contrastive Learning Regularization (HCLR-Net). Our method is built upon a distinctive hybrid contrastive learning regularization strategy that incorporates a unique methodology for constructing negative samples. This approach enables the network to develop a more robust sample distribution. Notably, we utilize non-paired data for both positive and negative samples, with negative samples are innovatively reconstructed using local patch perturbations. This strategy overcomes the constraints of relying solely on paired data, boosting the model’s potential for generalization. The HCLR-Net also incorporates an Adaptive Hybrid Attention module and a Detail Repair Branch for effective feature extraction and texture detail restoration, respectively. Comprehensive experiments demonstrate the superiority of our method, which shows substantial improvements over several state-of-the-art methods in terms of quantitative metrics, significantly enhances the visual quality of underwater images, establishing its innovative and practical applicability. Our code is available at: https://github.com/zhoujingchun03/HCLR-Net. [ABSTRACT FROM AUTHOR]

We present a novel method for underwater image restoration, which combines a Comprehensive Imaging Formation Model with prior knowledge and unsupervised techniques. Our approach has two main components: depth map estimation using a Channel Intensity Prior (CIP) and backscatter elimination through Adaptive Dark Pixels (ADP). The CIP effectively mitigates issues caused by solid-colored objects and highlighted regions in underwater scenarios. The ADP, utilizing a dynamic depth conversion, addresses issues associated with narrow depth ranges and backscatter. Furthermore, an unsupervised method is employed to enhance the accuracy of monocular depth estimation and reduce artificial illumination influence. The final output is refined via color compensation and a blue-green channel color balance factor, delivering artifact-free images. Experimental results show that our approach outperforms state-of-the-art methods, demonstrating its efficacy in dealing with uneven lighting and diverse underwater environments. [ABSTRACT FROM AUTHOR]