Search

In this paper, machine learning (ML) techniques are applied at an early stage of Image Processing (IP). The learning procedures are usually applied from at least the image segmentation level, whereas, in this paper, this is done from a lower processing level: the edge detection level (ED). The main objective is to solve the edge detection problem through ML techniques. The proposed methodology is based on a classification of edges made pixel by pixel, but the predictors employed for the ML task include information about the pixel neighborhood and structures of connected pixels called edge segments. The Sobel operator is employed as input. Making use of 50 images that belong to the Berkeley Computer Vision data set, the average performance of the validation sets when employing our Neural Networks method reached an F-measure significatively higher than with the Sobel operator. The experiment results show that our post-processing technique is a promising new approach for ED. [ABSTRACT FROM AUTHOR]

Multilevel thresholding for image segmentation is one of the crucial techniques in image processing. Even though numerous methods have been proposed in literature, it is still a challenge for the existing methods to produce steady satisfactory thresholds at manageable computational cost in segmenting images with various unknown properties. In this paper, a non-revisiting quantum-behaved particle swarm optimization (NrQPSO) algorithm is proposed to find the optimal multilevel thresholds for gray-level images. The proposed NrQPSO uses the non-revisiting scheme to avoid the re-evaluation of the evaluated solution candidates. To reduce the unnecessary computation cost, the NrQPSO provides an automatic stopping mechanism which is capable of gauging the progress of exploration and stops the algorithm rationally in a natural manner. For further improving the computation efficiency, the NrQPSO employs a meticulous solution search method to overcome the drawback of the existing QPSO algorithms using the original search methods. Performance of the NrQPSO is tested on the Berkeley segmentation data set. The experimental results have demonstrated that the NrQPSO can outperform the other state-of-the-art population-based thresholding methods in terms of efficiency, effectiveness and robustness; thus, the NrQPSO can be applied in real-time massive image processing. [ABSTRACT FROM AUTHOR]

In this paper we introduce a method for color image segmentation by computing automatically the number of clusters the data, pixels, are divided into using fuzzy c-means. In several works the number of clusters is defined by the user. In other ones the number of clusters is computed by obtaining the number of dominant colors, which is determined with unsupervised neural networks (NN) trained with the image's colors; the number of dominant colors is defined by the number of the most activated neurons. The drawbacks with this approach are as follows: (1) The NN must be trained every time a new image is given and (2) despite employing different color spaces, the intensity data of colors are used, so the undesired effects of non-uniform illumination may affect computing the number of dominant colors. Our proposal consists in processing the images with an unsupervised NN trained previously with chromaticity samples of different colors; the number of the neurons with the highest activation occurrences defines the number of clusters the image is segmented. By training the NN with chromatic data of colors it can be employed to process any image without training it again, and our approach is, to some extent, robust to non-uniform illumination. We perform experiments with the images of the Berkeley segmentation database, using competitive NN and self-organizing maps; we compute and compare the quantitative evaluation of the segmented images obtained with related works using the probabilistic random index and variation of information metrics. [ABSTRACT FROM AUTHOR]

• An efficient method for image thresholding is proposed. • The Harris hawk optimizer is used for image segmentation. • The Harris hawk optimizer is tested over a multidimensional real problem. • The quality of the segmentation results is higher than other algorithms. • The performance of the algorithm is tested on digital mammograms. Segmentation is a crucial phase in image processing because it simplifies the representation of an image and facilitates its analysis. The multilevel thresholding method is more efficient for segmenting digital mammograms compared to the classic bi-level thresholding since it uses a higher number of intensities to represent different regions in the image. In the literature, there are different techniques for multilevel segmentation; however, most of these approaches do not obtain good segmented images. In addition, they are computationally expensive. Recently, statistical criteria such as Otsu, Kapur, and cross-entropy have been utilized in combination with evolutionary and swarm-based strategies to investigate the optimal threshold values for multilevel segmentation. In this paper, an efficient methodology for multilevel segmentation is proposed using the Harris Hawks Optimization (HHO) algorithm and the minimum cross-entropy as a fitness function. To substantiate the results and effectiveness of the HHO-based method, it has been tested over a benchmark set of reference images, with the Berkeley segmentation database, and with medical images of digital mammography. The proposed HHO-based solver is verified based on other comparable optimizers and two machine learning algorithms K -means and the Fuzzy IterAg. The comparisons were performed based on three groups. This first one is to provide evidence of the optimization capabilities of the HHO using the Wilcoxon test, and the second is to verify segmented image quality using the PSNR, SSIM, and FSIM metrics. Then, the third way is to verify the segmented image comparing it with the ground-truth through the metrics PRI, GCE, and VoI. The experimental results, which are validated by statistical analysis, show that the introduced method produces efficient and reliable results in terms of quality, consistency, and accuracy in comparison with the other methods. This HHO-based method presents an improvement over other segmentation approaches that are currently used in the literature. [ABSTRACT FROM AUTHOR]

• Pattern complexity can be used to divide a task and improve learning efficiency. • Weighted mixtures of locally-learned networks boost generalization performance. • Training of networks is completely and naturally parallelizable by our strategy. • Adaptive ensembles of local small-scale networks beat a single large-scale one. Image denoising is a fundamental task in computer vision and image processing domain. In recent years, the task has been tackled with deep neural networks by learning the patterns of noises and image patches. However, because of the high diversity of natural image patches and noise distributions, a huge network with a large amount of training data is necessary to obtain a state-of-the-art performance. In this paper, we propose a novel ensemble strategy of exploiting multiple deep neural networks for efficient deep learning of image denoising. We divide the task of image denoising into several local subtasks according to the complexity of clean image patches and conquer each subtask using a network trained on its local space. Then, we combine the local subtasks at test time by applying the set of networks to each noisy patch as a weighted mixture, where the mixture weights are determined by the likelihood of each network for each noisy patch. Our methodology of using locally-learned networks based on patch complexity effectively decreases the diversity of image patches at each single network, and their adaptively-weighted mixture to the input combines the local subtasks efficiently. Extensive experimental results on Berkeley segmentation dataset and standard test images demonstrate that our strategy significantly boosts denoising performance in comparison to using a single network of the same total capacity. Furthermore, our method outperforms previous methods with much smaller training samples and trainable parameters, and so with much reduced time complexity both in training and running. [ABSTRACT FROM AUTHOR]