Similarity Of Neighboring Pixels Research Articles

Coarse2Fine: Local Consistency Aware Re-prediction for Weakly Supervised Object Localization

Weakly supervised object localization aims to localize objects of interest by using only image-level labels. Existing methods generally segment activation map by threshold to obtain mask and generate bounding box. However, the activation map is locally inconsistent, i.e., similar neighboring pixels of the same object are not equally activated, which leads to the blurred boundary issue: the localization result is sensitive to the threshold, and the mask obtained directly from the activation map loses the fine contours of the object, making it difficult to obtain a tight bounding box. In this paper, we introduce the Local Consistency Aware Re-prediction (LCAR) framework, which aims to recover the complete fine object mask from locally inconsistent activation map and hence obtain a tight bounding box. To this end, we propose the self-guided re-prediction module (SGRM), which employs a novel superpixel aggregation network to replace the post-processing of threshold segmentation. In order to derive more reliable pseudo label from the activation map to supervise the SGRM, we further design an affinity refinement module (ARM) that utilizes the original image feature to better align the activation map with the image appearance, and design a self-distillation CAM (SD-CAM) to alleviate the locator dependence on saliency. Experiments demonstrate that our LCAR outperforms the state-of-the-art on both the CUB-200-2011 and ILSVRC datasets, achieving 95.89% and 70.72% of GT-Know localization accuracy, respectively.

Spatial clustering using neighborhood for multispectral images

Spatial data mining discovers patterns and knowledge in spatial data. The geospatial data analysis plays a decisive role in framing essential policies related to the environment at the national as well as global level. It helps in the prediction of weather, studying the impact of climate change, the effect of global warming, forest fire, deforestation, and other causes of changes in nature. Clustering is a process of assigning similar objects to a group and dissimilar objects to different groups. We proposed a clustering algorithm, spatial clustering using neighborhood (SCN), in which similar neighboring pixels groups together to form a cluster. It is a two-step process: first, identification of clusters of similar neighboring pixels and second, merging of spatially separated similar clusters. Experiments with the implementation of the proposed approach have been carried out on three Landsat 5 thematic mapper images of Delhi, Chilika lake, and Kolkata region. In terms of normalized cluster quality parameters, SCN produced better clusters than other discussed algorithms in the literature.

Density Peak Covariance Matrix for Feature Extraction of Hyperspectral Image

The clustering methods have a good application in many aspects, in which the density peak (DP) clustering can effectively cluster similar neighboring pixels so that the features can be extracted well for hyperspectral images (HSIs) classification. In this work, a DP based covariance matrix (DPCM) method is proposed for the feature extraction of HSIs, which not only can effectively extract features but also can reduce the within-class variations and the between-class interference. The proposed method consists of the following steps: First, maximum noise fraction is employed on the original HSI to reduce the computational complexity and eliminate noise. Second, the local densities of the sample are calculated by the DP clustering. Therefore, a reconstructed image can be obtained in which each pixel has a density feature vector. Then, the covariance matrix between each density pixel in the density map is calculated. Last, the extracted covariance matrices are fed back to the support vector machine based on the logarithm Euclidean kernel for label assignment. Experiments are conducted on the Indian pine data set, in which each of the five randomly selected marker data are selected as the training sample. The experimental results show that the method can effectively improve the classification accuracy and is superior to other classification methods.

A robust adaptive spatial and temporal image fusion model for complex land surface changes

Spatial and temporal satellite image fusion (STIF) has provided a feasible alternative for generating imagery with both high spatial and temporal resolution, thus expanding the applications of existing satellite sensors. However, a critical challenge confronting the further development of STIF is to systematically and robustly address complex land surface changes, which include land cover changes without shape changes (e.g., crop rotation) and land cover changes with shape changes (e.g., urban expansion), in addition to conventional land surface changes (e.g., phenological changes of vegetation). This paper presents the Robust Adaptive Spatial and Temporal Fusion Model (RASTFM) to tackle this challenge with one prior pair of MODIS-Landsat images. In RASTFM, land surface changes are reorganized into non-shape changes (including phenological changes and land cover changes without shape changes) and shape changes (i.e., land cover changes with shape changes), which are handled differently. However, both non-shape changes and shape changes are predicted through a Non-Local Linear Regression (NL-LR) of the subject pixel's similar neighbors. A regression based high-pass modulation is also performed as a post-processing step to improve both the spatial details and spectral fidelity of the predicted Landsat image. Unlike other STIF models (e.g., the Spatial and Temporal Adaptive Reflectance Fusion Model, STARFM), RASTFM can find similar neighboring pixels more precisely through a non-local searching strategy and derives the weights of similar neighbors more rigorously via a linear regression model. As both non-shape and shape changes are treated based on the regression of similar neighboring pixels, the land surface changes are processed in a unified manner. Experiments that use one simulated and three actual MODIS-Landsat datasets featured by different types of land surface changes were conducted to demonstrate the performance of RASTFM. Comparisons with the state-of-the-art STIF models, including weighted function, unmixing and dictionary-learning methods, show that NL-LR based RASTFM can capture the land surface changes in various landscapes more accurately and robustly in a unified manner, thereby facilitating the continuous and detailed monitoring of complex and diverse land surface dynamics.

Hyperspectral image filtering with adaptive manifold for classification

Hyperspectral image (HSI) is a three-dimensional data cube containing two spatial information dimensions and one spectral information dimension. The spectral vectors of different classes may have similar tendency and value that may bring about negative influences on classification. It is, therefore, important to introduce signal preprocessing techniques in the spatial domain to improve classification accuracy of HSIs. Assuming that local pixels in HSI have some correlations with each other, this paper proposes a spatial filtering model based on adaptive manifold (AM) for HSI. The AM for spatial filtering emphasizes the similar neighboring pixels and is robust to resist the noisy points with fast speed. The rich information in the filtered data is effective for improving the performance of the subsequent classification. The filtered data are classified by an extreme learning machine (ELM). The experimental results indicate that the framework built based on AM and ELM provides competitive performance. Specifically, by classifying the filtered data, the average accuracy of ELM can be improved as high as 30.54%, while performing tens to hundreds times faster than those state-of-the-art classifiers.

Fast Low-Rank Decomposition Model-Based Hyperspectral Image Classification Method

In hyperspectral image classification, jointly using the pixels in an image patch can generally improve the performance. Recently, a new hyperspectral image classification method, which is based on low-rank decomposition model, was proposed by Chen et al. Although this algorithm can achieve state-of-the-art performance and outperform many contemporary classification techniques by jointly classifying both the central pixel and its similar neighboring pixels in the image patch, its computational load is heavy due to its pixelwise processing scheme, which limits its possible applications in practice. In this letter, we proposed a fast groupwise version of this method. In this improved method, the low-rank decomposition model is first used to partition the hyperspectral scene into many groups of similar pixels. Then, the pixels in each similar group would be classified jointly and assigned the same class label. The proposed groupwise classification method can achieve similar performance with the approach described by Chen et al. but with much reduced computational load. Our results demonstrate that when more observations are simultaneously used, the performance of the jointly sparse regression model would increase, and when the number of observations reaches a certain amount, the performance would no longer increase significantly. Our experimental results also demonstrate that, when the jointly sparse regression model is used, the quality is preferred to the quantity of the observations after the number of observations has reached a certain amount.

Low-Rank Decomposition Model for Adaptive Identification of Similar Neighboring Pixels in Hyperspectral Images

In many hyperspectral image processing tasks such as classification, unmixing, and target detection, jointly using the pixels in an image patch can generally improve the performance. In this letter, we propose using a low-rank decomposition model to analyze the image patch. The image patch is decomposed into the sum of a low-rank matrix, a sparse matrix, and a bounded matrix using a convex optimization technique. In the obtained low-rank matrix, the pixels that are similar to the central pixel in the image patch would be codirectional to it. We applied the proposed model in the collaborative hyperspectral image classification to evaluate its performance. Experimental results on a real hyperspectral scene demonstrate that using only the similar neighboring pixels in the collaborative hyperspectral image classification can effectively improve the performance, and the classification performance is not sensitive to the image patch size.

A Local Similarity Pattern for Removal of Random Valued Impulse Noise

This letter presents a new filtering scheme based on local similarity pattern within the local window for removing random valued impulse noise. A pixel to be considered as an original pixel, it should have ample numbers of similar neighboring pixels in a local window. The neighbors are divided into 2 subtypes: smooth similar neighbors and edge similar neighbors according to the different criterion. Extensive simulations show that the proposed filter provides better performance than many of the existing filters. In particular, the thresholds are adaptive to diverse image types at different noise rate and the computational complexity is very low.

Similar neighbor criterion for impulse noise removal in images

In this paper, a new method is proposed for removing and restoring random-valued impulse noise in images. This approach is based on a similar neighbor criterion, in which any pixel to be considered as an original pixel it should have sufficient numbers of similar neighboring pixels in a set of filtering windows. Compared with other well known methods in the literature, this technique achieves superior performance in restoring heavily corrupted noisy images. Furthermore, it has low computational complexity, and equally effective in restoring corrupted color and gray-level images.

Scene segmentation using a combined criterion of motion and intensity

An image segmentation algorithm for segmentation-based video sequence coding is presented. The method uses a new criterion based on similarities in both motion and brightness. Brightness and motion information are incorporated in a single segmentation procedure. The actual segmentation is accomplished using a region-growing technique based on the watershed algorithm. It comprises the processes of extracting the initial regions called seeds and their growing by merging similar neighboring pixels. In addition, a tracking technique is used in subsequent frames to achieve a coherent segmentation through time. Simulation results show that the proposed method is effective in determining object boundaries not easily found using the static criterion alone.