نشریه ماشین بینایی و پردازش تصویر
سال نهم شماره 1 (بهار 1401)

Imbalanced image classification is one of the most important and difficult issues in data mining. With the inability of standard classification algorithms, Capsule neural networks (CapsNet) provide a good platform for designing imbalanced classification models by considering spatial communication of features, compared to other deep networks such as Convolutional Neural Networks (CNN). On the other hand, crack bifurcation in the surface cracks is one of the anomalies and minority categories in concrete structures that can be effective in the maintenance of concrete structures and cost management. Also, the surface crack image sets are suitable data for evaluating imbalanced classification due to their characteristics. Therefore, in this paper, a new architecture based on CapsNet is introduced to evaluate the imbalanced classification of surface crack images in the concrete structures. Examination and comparison of the proposed network with CNN in balanced and imbalanced image classification of surface cracks on 13,500 sets of collected images showed the superiority of the proposed network. Also, the proposed network showed a significant advantage compared to CNN in investigating the effect of reducing the number of training images on classification accuracy. This network performed balanced classification of surface cracks with 99.56% accuracy. Also, the proposed network has an accuracy of 80% up to the imbalance of theminority group to the 1:8 minority, which is very suitable compared to CNN.

In the traditional methods of analyzing minerals in thin sections, the boundaries of the minerals were manually separated and each section was labeled. This approach is expensive and requires high expertise and experience. Therefore, an automatic identification system is essential in this field. Such a system can increase the accuracy and reduce human error, cost and time of mineral identification. The aim of this study is to propose an automated identification system which uses image processing to identify and classify existing minerals.The main steps of the proposed method include collecting images from thin sections, segmentation, feature extraction and classification. After creating the image database, the JSEG algorithm is applied for segmentation. Then, the color and texture features in both RGB and HSI color spaces are extracted from each region and are sent to the classifier for classification, which labels each segment as a mineral. In this study, the efficiency of six different classifiers has been evaluated. According to the results, the Bagged Tree classifier has the highest accuracy of 95.52% and the lowest Mean Absolute Error of 0.04. Also, all classifiers have accuracies of over 93%, which indicates that the proposed feature extraction method is able to properly identify minerals.

One of the challenges of convolutional neural networks (CNNs), as the main tool of deep learning, is the large volume of some relevant models. CNNs, inspired form the brain, have millions of connections. Reducing the volume of these models is done by removing (pruning) the redundant connections of the model. Optimal Brain Damage (OBD) and Sparse Regularization are among the famous methods in this field. In this study, a deep learning model has been trained and the effect of reducing connections with the aforementioned methods on its performance has been investigated. As the proposed approach, by combining the OBD and regularization methods its redundant connections were pruned. The resulting model is a smaller model, which has less memory and computational load than the original model, and at the same time its performance is not less than the original model. The experimental results show that the hybrid approach can be more efficient than each of the methods, in the most tested datasets. In one dataset , with the proposed method, the number of connections were reduced by 76%, without sacrificing the efficiency of the model. This reduction in model size has decreased the processing time by 66 percent. The smaller the software model, the more likely it is to be used on weaker hardware, found everywhere, and web applications.

In the last few decades, a lot of research has been done in the field of OCR or optical character recognition. Optical character recognition is one of the ways to convert text images to editable text and recognize letters and words automatically. Recognizing textual and non-textual areas within a document is known as document layout analysis, and is one of the key steps in the process of converting a document image to editable text. Separating textual and non-textual areas within an image is one of the most effective possible preprocesses in optical character recognition systems. The lack of the same template on all pages, the presence of complex backgrounds, different kinds of noises, low quality, image rotation, and the existence of more than one text column prevent the correct recognition of areas containing text. Failure to correctly recognize areas containing text and, consequently, incorrect recognition of line coordinates will disrupt all subsequent parts of an optical character recognition system. In this research, a new method has been proposed to recognize textual areas within the image. The proposed method, using various methods and using a voting system among them, extracts the textual areas of the image. The proposed method has been trained and tested on a dataset with more than 950 images and reached 97.94% accuracy. The presented dataset in this article is open access.

In machine learning, transferring and generalizing the knowledge learned from one domain to another is one of the important and basic capabilities. Since supervised learning is not complete, the use of other methods, such as self-supervised learning methods, can be very helpful in domain generalization. In this paper, we present a method that, in addition to classify original images in order to learn data labels in a supervised process, attempts to classify images resulting from the application of discrete wavelet transform on the original images by generating pseudo-labels for them. This extra work as a self-supervision task can lead to learn useful features and a general image representation for images of different domains, which can greatly help to improve the problem of domain generalization. In the following, by combining self-supervised methods such as jigsaw puzzles and guessing the rotation angle with discrete wavelet transform, we show that this combination can improve the results for the domain generalization problem. In this paper, we used the well-known PACS, VLCS and office-Home datasets to perform experiments, and the results show that our proposed method can work better than advanced and state-of-the-art domain generalization methods.

In this paper, a novel method has been proposed to automatically produce 3D videos through amateur digital cameras that have been fixed with each other. Asynchrony of the videos (frames per second rates and start times), unavailability of the exact camera internal parameters, and the technical limitation of precise relative adjusting the stereo cameras could be known as the main challenges of producing 3D videos via the amateur cameras. In the proposed method, the videos acquired by the stereo camera are synchronized through the automatic matching between temporal indices. Then, the calm periods of videos (the times with zero relative velocity between camera and scenes) are detected to be used for selecting proper matched image frames. Proper matched frames are then applied for finding matched points via a geometrical constrained feature-based matching method. The matched points are used for self-calibration as well as relative parameters estimation of the stereo cameras. In the last step, epipolar resampling procedure has been used to generate normalized videos. Automatic and precise synchronization of the stereo videos as well as the proper generalization of the proposed approach in the different sample datasets has been seen in the evaluation processes. Spectator satisfaction in the depth perception of the 3D videos is another quality achievement of the proposed method.