نشریه ماشین بینایی و پردازش تصویر
سال دهم شماره 3 (پاییز 1402)

Image captioning is a new field of artificial intelligence which uses both natural language processing and machine learning methods. Recently, several researches conducted on Image captioning have led to the acceptable results in this field. Moreover, the need of the medical community for automatic interpretation of medical images, prompted researchers to apply this approach in the field of automatic interpretation of medical images. The medical images captioning is more challenging than the problem of image description because medical images are mostly gray and accompanied by noise, the existing data sets have significantly few samples; On the other hand, the provided interpretations are not structured and have variable lengths. In order to achieve the desired results, many algorithms have been proposed, among which methods based on deep learning have achieved proper results. This article introduces datasets, evaluation criteria and methods developed based on deep learning in the field of automatic interpretation of medical images in order to help understand the existing literature and highlight future directions in this field.

Coral reefs are an important part of the tropical shallow water ecosystem and their protection is very important. Classification of coral reef images includes three stages of image enhancement, feature extraction and classification. In this research, by focusing on the feature extraction a method for features extraction for classification of coral corals images is proposed. This method consists of two methods of local binary pattern variants. In addition, instead of using a large neighborhoods, a multi-scale neighborhood with different sizes is used. This method employed a fixed number of points with different size of neighborhoods. This increases the classification accuracy without exponentially increasing the features. By combining the CS_LBP symmetric binary method and the MRELBP median enhanced binary method, some features of the image are merged together, and the local features extracted by the CS_LBP method are reduced by half in each step. In this research, the accuracy of the proposed model has been evaluated on EILAT, EILAT2, RSMAS, and MLC-2008 coral reef image sets. Also a general textures such as CUReT, UIUC, and KTH_TIPS texture are used. The classification accuracy of the proposed method has increased in all recent data, while the number of features extracted is decreased.

Enlarging digital images is one of the image processing methods, which improves the image in the field of computer vision. Basically, this is used to zoom in on still and moving images that have passed the time of their capture and there is no access to the camera or scenes to zoom. In this article, a hierarchy is proposed to extract high-level features to solve the demarcation challenge between colors and self-learning attention blocks to reduce convolution operations. In the following, to optimize the network, random search and binary division are used to find optimal answers and meta-parameters. Using the mentioned search method, in addition to searching for network weights and networks, I can also search for the architecture of the structure, this action will automatically generate meta-parameters and optimize the network structure. To check the effectiveness of the proposed method, simulation results on the database in this area have been determined, which show that the proposed method is superior to other methods. According to the results obtained in architecture, by using the four-fold magnification mentioned in the four-story series block and using the attention block in the magnification section, the display-to-noise number was 32.66.

Nowadays, traffic has become a challenge for everyone. One of the ways to overcome this problem is to make cars more intelligent. In this research, perception enhancement of the environment has been considered by using the data fusion of ultrasonic data and stereo vision. In this research, the researchers have suggested the cut stereo vision method by ultrasonic data in such a way that the accuracy and speed of obstacle detection will increase simultaneously in smart cars. Therefore, in addition to the similarity of the light intensity in the matching window, the neighboring pixels' depth has been used in a way to achieve the goals mentioned above without exponentially increasing the computational load. So, two types of compliance windows have been defined. One window is similar to the cut stereo method, and the other is called the inference window. By using the proposed inference window based on cut stereo vision, it is possible to apply the influence of the depth of neighboring pixels at a sufficient and effective level, which is reduced 61.57% error compared to the previous study. Furthermore, according to the suggested way of implementation on the graphic card, the obstacle detection speed has improved by 43.93% compared to the previous parallel implementation. The mentioned improvements make it possible to detect the environment in each meter of a car's movement if its speed is 178.1 km/h rather than 100.8km/h in the previous study. So current study produces smart cars more reliable.

Identifying objects, tracking objects and predicting time series are among the basic challenges in machine vision. Deep learning has taken great steps in solving these challenges; But for many problems, satisfactory solutions that have useful applications in reality and can be used have not yet been found. In this issue, we are facing two challenges of tracking and identifying objects, and to solve this problem, it is proposed to find limiting tubes for the movement of objects in the space-time domain. Usually, object tracking and object detection are considered as two separate processes, which have been greatly improved through deep learning for 2D images. Object tracking by object detection requires that the object is successfully detected in the first frame and in all subsequent frames, and thus, by associating the results obtained from object detection, we performed the tracking operation by the TPN pipeline. The operation of identifying objects and tracking objects through a single network is still challenging and debatable. In this paper, a network structure is proposed that was able to identify a moving and moving object that was enclosed, using R-CNN Faster. In this network, we replaced TPN with RPN, and this led to better object identification and improved tracking. In this method, we tracked objects using object detection operations.

Video surveillance cameras can be used as a powerful tool for automating the detection of various situations and helping to make appropriate decisions in order to increase the level of security and protection. One of the most important applications of video surveillance systems is the detection of abandoned objects such as abandoned luggage is to prevent dangerous bombings and other cases. In this regard, in this article, a two-stage model based on deep learning has been introduced to detect abandoned objects. The purpose of the first stage is to detect all the objects in the scene and the second stage is to classify the abandoned objects. In the first step, the Gaussian mixture model (GMM) is used to model the background and detect stationary objects. In the second step, a combination of convolutional neural network (CNN) and the AdaBoost algorithm is used to identify the abandoned objects among all the extracted images. Based on the results of the evaluations, the proposed model has a higher accuracy in detecting abandoned objects than the basic methods.