Set Of Test Images Research Articles

Single Banana Appearance Grading with Ppyolo-Banana

Highlights An inspection method for grading the appearance of bananas using a target detection algorithm is proposed—by calculating the number and area ratio of different defective areas as a discriminating criterion. The Mish activation function makes the network easier to optimize and improves generalization performance. CustomPAN adds an attention mechanism, optimized for better multi-feature fusion. Optimized loss function in regression task with DIoULoss. Abstract. With the development of the fruit individual packaging industry, the appearance quality of individually packaged fruits has put forward higher requirements. Due to the dense and uneven defects on the surface of bananas, the existing detection algorithms are prone to the problem of unrecognizable or degraded recognition accuracy. In this article, we propose an efficient banana surface defect detection model, the PPYOLO-Banana model. PPYOLO-Banana is based on the PPYOLOE+-m model with improved model structure and loss function, and the optimized CustomPAN can get more multi-level features, and compared with the original network PPYOLOE+-m model, the algorithm significantly improves the accuracy, with an average accuracy improvement of 2.2% (1.3% for the original image test set). mAP of PPYOLO-Banana is 97.0% (96.1% for the original image test set), which is 14.3% higher than the PPYOLOE model, and 10.9%, 8.9%, 8.9%, and 8.1% higher than the YOLOX, YOLOX-tiny, YOLOv5, and YOLOV4 models, respectively. The detection speed of the PPYOLO-Banana model is 17.71 frames per second, which is 2.95, 2.10, 1.90, and 0.98 times higher than that of YOLOv3, YOLOv4, YOLOX, and YOLOX-tiny, respectively. The results show that the proposed PPYOLO-Banana model achieves a balance between accuracy and speed in recognizing banana surface defects, improves the quality detection capability of individually packed fruits, it can effectively grade the quality of banana appearance, and has good potential to become an intelligent sorting machine. Keywords: Banana defect recognition, Banana appearance grading, CustomPAN, DIoULoss, PPYOLOE+.

Counting microalgae cultures with a stereo microscope and a cell phone using deep learning online resources

BackgroundThis work presents an experience done to evaluate the number of very small objects in the field of view of a stereo microscope, which are usually counted by direct observation, with or without the use of grids as visual aids. We intend to show that deep learning recent algorithms like YOLO v5 are adequate to use in the evaluation of the number of objects presented, which can easily reach the 1000 s. This kind of algorithm is open-source software, requiring a minimum of skills to install and run on a regular laptop. We further intend to show that the robustness of these kinds of approaches using convolutional neural networks allowed for the use of images of less quality, such as the images acquired with a cell phone.ResultsThe results of training the algorithm and counting microalgae in cell phone images were assessed through human curation in a set of test images and showed a high correlation, showing good precision and accuracy in detections.ConclusionsThis is a low-cost alternative available worldwide to many more facilities than expensive cameras and high-maintenance rigid set-ups, along with software packages with a slow learning curve, therefore enlarging the scope of this technique to areas of knowledge where the conditions of laboratory and human work are a limiting factor.

Automatic identifying and counting blood cells in smear images by using single shot detector and Taguchi method

Researchers have tried to identify and count different blood cells in microscopic smear images by using deep learning methods of artificial intelligence to solve the highly time-consuming problem. The three types of blood cells are platelets, red blood cells, and white blood cells. This study used the Resnet50 network as a backbone network of the single shot detector (SSD) for automatically identifying and counting different blood cells and, meanwhile, proposed a systematic method to find a better combination of algorithm hyperparameters of the Resnet50 network for promoting accuracy for identifying and counting blood cells. The Resnet50 backbone network of the SSD with its optimized algorithm hyperparameters, which is called the Resnet50-SSD model, was developed to enhance the feature extraction ability for identifying and counting blood cells. Furthermore, the algorithm hyperparameters of Resnet50 backbone networks of the SSD were optimized by the Taguchi experimental method for promoting detection accuracy of the Resnet50-SSD model. The experimental result shows that the detection accuracy of the Resnet50-SSD model with 512 × 512 × 3 input images was better than that of the Resnet50-SSD model with 300 × 300 × 3 input images on the test set of blood cells images. Additionally, the detection accuracy of the Resnet50-SSD model using the combination of algorithm hyperparameters got by the Taguchi method was better than that of the Resnet50-SSD model using the combination of algorithm hyperparameters given by the Matlab example. In blood cell images acquired from the BCCD dataset, the proposed Resnet50-SSD model had higher accuracy in identifying and counting blood cells, especially white blood cells and red blood cells.

Deep learning classification of urinary sediment crystals with optimal parameter tuning

The examination of urinary sediment crystals, the sedimentary components of urine, is useful in screening tests, and is always performed in medical examinations. The examination of urinary sediment crystals is typically done by classifying them under a microscope. Although automated analyzers are commercially available, manual classification is required, which is time-consuming and varies depending on the technologist performing the test and the laboratory. A set of test images was created, consisting of training, validation, and test images. The training images were transformed and augmented using various methods. The test images were classified to determine the patterns that could be correctly classified. Convolutional neural networks were used for training. Furthermore, we also considered the case where the crystal subcategories were not treated as separate. Learning with all parameters except the random cropping parameter showed the highest accuracy value. Treating the subcategories together or separately did not seem to affect the accuracy value. The accuracy of the best pattern was 0.918. When matched to a real-world case, the percentage of correct answers was 88%. Although the number of images was limited, good results were obtained in the classification of crystal images with optimal parameter tuning. The parameter optimization performed in this study can be used as a reference for future studies, with the goal of image classification by deep learning in clinical practice.

A Pixel Based Method for Image Compression

The basic solution to overcome difficult issues related to huge size of digital images is to recruited image compression techniques to reduce images size for efficient storage and fast transmission. In this paper, a new scheme of pixel base technique is proposed for grayscale image compression that implicitly utilize hybrid techniques of spatial modelling base technique of minimum residual along with transformed technique of Discrete Wavelet Transform (DWT) that also impels mixed between lossless and lossy techniques to ensure highly performance in terms of compression ratio and quality. The proposed technique has been applied on a set of standard test images and the results obtained are significantly encourage compared with Joint Photographic Experts Group (JPEG).

Classification of Pomegranate Leaves Diseases by Image Processing and Machine Learning Techniques

A decrease in agricultural yield has been caused by uneven climatic conditions. Global agricultural economies are affected by this. Furthermore, if the crops are infected by any disease, the situation becomes even worse. In addition, farmers have to increase their yield in order to cope with a growing population. In order to detect and prevent various diseases from affecting crops, modern agricultural techniques, and systems are needed. One of the most profitable fruits on the market is the pomegranate, a fruit that produces an astounding yield in a number of Asian nations. Unfortunately, due to a number of conditions, the plants are infected with several diseases that destroy the entire crop leaving a terribly low yield of products. A key challenge in agricultural image processing is the recognition of disease in plants. To identify and classify diseases on pomegranate plants precisely, this framework uses image processing techniques such as acquisition, resizing, enhancement, segmentation, ROI extraction, and feature extraction to process images. The framework is implemented using a training set and a test set of images associated with pomegranate leaf disease. Image enhancement and image segmentation are usually used during the implementation process in order to identify ROIs and features. The use of machine learning (ML) may be a viable way to solve this problem. The problems are solved using ML models like linear discriminant analysis (LDA), K-Nearest Neighbor, Naive Bayes (NB), and ensembles (LDA–NB). By combining LDA with NB, the framework can achieve an accuracy of 96.49% for classifying diseases on pomegranate leaves.

АНАЛИЗ ПОДСТИЛАЮЩЕЙ ПОВЕРХНОСТИ ПРИ ФОРМИРОВАНИИ ИЗОБРАЖЕНИЯ В РЕЖИМЕ ДОПЛЕРОВСКОГО ОБУЖЕНИЯ ЛУЧА

Radar based on real beam scanning is widely used in both civil and military spheres. However,it is difficult to realize high azimuth resolution of a stationary platform or a platform with nonuniformmotion using conventional signal processing algorithms. Doppler beam sharpening (DBS)technology is a combination of high resolution and real-time performance compared to SyntheticAperture Radar (SAR) technology, which uses the Doppler shift between echoes from objects on theunderlying surface along the azimuth direction, caused by the movement of the radar platform. Unfortunately,the traditional DBS imaging algorithm, which constructs a Doppler filter using an FFT,has a low azimuth resolution and a high level of side lobes, which limits further improvement in azimuthresolution. In the article, the algorithm for constructing a map of the underlying surface in thedirection of movement of the radar carrier based on the DBS was studied and the map image wasanalyzed using the Fourier transform. A three-dimensional view of the map of the underlying surfaceis shown with the distribution of values in the images. The subject of the study is the method and algorithmfor constructing a map of the underlying surface in the Doppler beam sharpening mode andidentifying chain structures based on the analysis of the Fourier transform. The object of the study is a set of test images of the terrain map. The result of the study is the development of an algorithm forconstructing a map in order to identify chain structures on the underlying surface. The novelty of thework is an algorithm that allows you to build a map of the underlying surface based on the DBS,taking into account the blind zone in the direction of movement of the radar carrier. The results obtainedalso make it possible to reveal chain structures in the region of interest. The possibility ofestimating the periodicity of image elements using the Fourier transform has been tested. As a resultof solving the tasks set, the following conclusions can be drawn: – an algorithm has been developedfor constructing a map of the underlying surface based on DBS with image correction in the directionof movement of the radar carrier. – analysis of the results of the study showed that the proposed algorithmallows you to identify chain structures on the map.

Histogram modification based pansharpening by using differential evolution algorithm

SUMMARYOptic geo‐observer satellites acquire multispectral images (MSI) with low spatial resolution and panchromatic images (PAN) with high spatial resolution, due to communication band‐limits and the necessity of using on‐board energy efficiently. Histogram modification is an image processing method that alters the intensity values of an image to improve contrast. The Histogram Modification based Pansharpening (HMP) method has been proposed in this article for injecting high‐resolution spatial information of PAN into MSI with limited color degeneration to generate a pansharpened image (PSI). The differential evolution algorithm (DE) has been used to optimize the internal parameters of the HMP. HMP uses three different blind‐image quality metrics to avoid from artificial spectral and spatial defects generally appeared in PSI. The quality of PSI generated by HMP has been statistically compared to the PSIs produced by 17 different pansharpening methods by using 10 different image quality metrics. In the tests performed, three different test image sets acquired by geo‐observation satellites have been used. Test image sets are covering the areas with different spatial and spectral characteristics. HMP outperforms the other state‐of‐the‐art pansharpening methods used in this article, according to the statistical analysis results. HMP is highly effective at reliably preserving spectral and spatial fidelity.

COMPARATIVE INVESTIGATION OF NEURAL AND LOCALLY BINARY ALGORITHMS FOR IMAGE IDENTIFICATION OF PLANKTON POPULATIONS

The work is devoted to the «neural network-lbp» method of processing satellite images of multispectral water coastal systems for identification of phytoplankton populations of spotted structure: determination of their boundaries, distribution of color gradations and, based on this, determination of the distribution of phytoplankton concentrations inside the spots and the location of the center of mass. The efficiency and effectiveness of the proposed neural network-lbp method is investigated in comparison with the method based on the use of a three-layer neural network. For the analysis, a test set of images is involved – flat shapes with rather complex boundaries, which will allow us to quantify the quality of the algorithms being compared. The results of the work show an increase in recognition accuracy by 1.5-3% when using the proposed method.

Automatic detection of crohn disease in wireless capsule endoscopic images using a deep convolutional neural network

The diagnosis of Crohn’s disease (CD) in the small bowel is generally performed by observing a very large number of images captured by capsule endoscopy (CE). This diagnostic technique entails a heavy workload for the specialists in terms of time spent reviewing the images. This paper presents a convolutional neural network capable of classifying the CE images to identify those ones affected by lesions indicative of the disease. The architecture of the proposed network was custom designed to solve this image classification problem. This allowed different design decisions to be made with the aim of improving its performance in terms of accuracy and processing speed compared to other state-of-the-art deep-learning-based reference architectures. The experimentation was carried out on a set of 15,972 images extracted from 31 CE videos of patients affected by CD, 7,986 of which showed lesions associated with the disease. The training, validation/selection and evaluation of the network was performed on 70%, 10% and 20% of the total images, respectively. The ROC curve obtained on the test image set has an area greater than 0.997, with points in a 95-99% sensitivity range associated with specificities of 99-96%. These figures are higher than those achieved by EfficientNet-B5, VGG-16, Xception or ResNet networks which also require an average processing time per image significantly higher than the one needed in the proposed architecture. Therefore, the network outlined in this paper is proving to be sufficiently promising to be considered for integration into tools used by specialists in their diagnosis of CD. In the sample of images analysed, the network was able to detect 99% of the images with lesions, filtering out for specialist review 96% of those with no signs of disease.

Classification of tropical cyclone containing images using a convolutional neural network: performance and sensitivity to the learning dataset

Abstract. Tropical cyclones (TCs) are one of the most devastating natural disasters, which justifies monitoring and prediction on short and long timescales in the context of a changing climate. In this study, we have adapted and tested a convolutional neural network (CNN) for the classification of reanalysis outputs according to the presence or absence of TCs. This study compares the performance and sensitivity of a CNN to the learning dataset. For this purpose, we chose two meteorological reanalysis, ERA5 and MERRA-2, and used a number of meteorological variables from them to form TC-containing and background images. The presence of TCs is labeled from the HURDAT2 dataset. Special attention was paid to the design of the background image set to make sure it samples similar locations and times to the TC-containing images. We have assessed the performance of the CNN using accuracy but also the more objective AUC and AUPRC metrics. Many failed classifications can be explained by the meteorological context, such as a situation with cyclonic activity but not yet classified as TCs by HURDAT2. We also tested the impact of spatial interpolation and of “mixing and matching” the training and test image sets on the performance of the CNN. We showed that applying an ERA5-trained CNN to MERRA-2 images works better than applying a MERRA-2-trained CNN to ERA5 images.

Synthesis of isognostic text images for visual performance evaluation

Subject of study. Principles and technologies for synthesis of test letter charts and standardization of measurements of reading parameters for evaluation of visual performance in the reading process, which is the most important type of human visual activity, are the subjects of the study. Aim of study. The study was aimed at developing a synthesis algorithm for textual isognostic images for quantitatively evaluating reading parameters as indicators of the state of visual performance that ensure the correlation between optical and semantic properties based on optical and optometric standards. Methods of study. A set of test images in the form of texts was created using digital methods for image synthesis. The text images were standardized for measuring reading parameters using optometric methods. Main results. A set of texts was synthesized and standardized considering the complexity and number of symbols in words, the frequency of occurrence and average width of Russian letters, the lengths of sentences, the angular sizes of characters, and the contrast. The text was standardized according to the angular visual acuity and conventional optometric standard ISO 8596, 2009. The synthesized test images comprised sets of text fragments of similar numbers of characters but different angular sizes. The sizes of the characters were correlated with the threshold angular values characterizing visual acuity according to optometric standards. The maximum reading speed, critical print size, and reading acuity were measured in 29 healthy adult test subjects to assess the reproducibility of results obtained using the synthesized test images. The measurements were performed twice for each test subject using two different charts synthesized using the developed method. No statistically significant difference was observed between the measurement results, which indicated good reproducibility of the results obtained using this test. Practical significance. The proposed algorithm for synthesis of standardized texts for assessing reading parameters is the simplest general method for a broad application in routine effectiveness evaluation of the therapy of macular zone pathology and lens fitting for short working distances.

More slices, less truth: effects of different test-set design strategies for magnetic resonance image classification

AimTo assess the effects of different test-set design strategies for magnetic resonance (MR) image classification using deep learning.MethodsError rates in 10 experimental settings were assessed. The performance of pretrained models and data augmentation were examined as possible contributing factors.ResultsError rates in experimental settings using MR images of different patients for training and test sets were ten times higher than those in experimental settings using MR images of the same patients (four disease groups with whole-chest images, 46.80% vs 2.06%; four disease groups without whole-chest images, 49.09% vs 1.29%; sex classification with whole-chest images, 16.02% vs 0.96%; and sex classification without whole-chest images, 23.56% vs 0.30%). Error rates were higher when data augmentation was applied to settings that used MR images of different patients for training and test sets.ConclusionWhen deep learning is applied to MR image classification, training and test sets should consist of MR images of different patients. Models built on training and test sets consisting of images of the same patients yield optimistic error rates and lead to wrong conclusions. MR images of neighboring slices are so similar that they cause data leakage effect.

Contrast stretching based pansharpening by using weighted differential evolution algorithm

Pansharpening techniques were developed to generate a super-resolution multispectral pansharpened image, PI, with the combination of a multispectral image carrying high-resolution spectral information with a panchromatic image carrying high resolution spatial information. For using energy resources and communication bandwidth efficiently, Earth Observation Satellites acquire multispectral images with lower spatial resolution when compared to panchromatic images. Contrast Stretching alters the range and statistical distribution of pixel values of an image to facilitate perception of image features and can be used to match histograms of distinct images. The Contrast Stretching Based Pansharpening method, CSP, has been introduced in this paper. CSP considers the pansharpening as a rescaling-based pixel-level image fusion problem in spatial domain. CSP uses the contrast stretching to generate two modified-multispectral images and one modified-panchromatic image, which are used to compute pansharpened image. The Weighted Differential Evolution Algorithm has been used to optimize the numerical values of internal parameters of CSP. The successes of CSP and 17 different pansharpening methods have been statistically compared by using three Test Image sets with different characteristics. Ten different image quality measures have been used in accuracy assessment analysis of the PIs generated by the related methods used in the Experiments. Statistical analysis exposed that CSP generates more pleasing PIs both quantitatively and qualitatively compared to the comparison methods employed in this paper.

Artificial Intelligence System for Automatic Quantitative Analysis and Radiology Reporting of Leg Length Radiographs.

Leg length discrepancies are common orthopedic problems with the potential for poor functional outcomes. These are frequently assessed using bilateral leg length radiographs. The objective was to determine whether an artificial intelligence (AI)-based image analysis system can accurately interpret long leg length radiographic images. We built an end-to-end system to analyze leg length radiographs and generate reports like radiologists, which involves measurement of lengths (femur, tibia, entire leg) and angles (mechanical axis and pelvic tilt), describes presence and location of orthopedic hardware, and reports laterality discrepancies. After IRB approval, a dataset of 1,726 extremities (863 images) from consecutive examinations at a tertiary referral center was retrospectively acquired and partitioned into train/validation and test sets. The training set was annotated and used to train a fasterRCNN-ResNet101 object detection convolutional neural network. A second-stage classifier using a EfficientNet-D0 model was trained to recognize the presence or absence of hardware within extracted joint image patches. The system was deployed in a custom web application that generated a preliminary radiology report. Performance of the system was evaluated using a holdout 220 image test set, annotated by 3 musculoskeletal fellowship trained radiologists. At the object detection level, the system demonstrated a recall of 0.98 and precision of 0.96 in detecting anatomic landmarks. Correlation coefficients between radiologist and AI-generated measurements for femur, tibia, and whole-leg lengths were > 0.99, with mean error of < 1%. Correlation coefficients for mechanical axis angle and pelvic tilt were 0.98 and 0.86, respectively, with mean absolute error of < 1°. AI hardware detection demonstrated an accuracy of 99.8%. Automatic quantitative and qualitative analysis of leg length radiographs using deep learning is feasible and holds potential in improving radiologist workflow.

The incremental online k-means clustering algorithm and its application to color quantization

Color quantization is a common image processing operation with various applications in computer graphics, image processing, and computer vision. Color quantization is essentially a large-scale combinatorial optimization problem. Many clustering algorithms, both of hierarchical and partitional types, have been applied to this problem since the 1980s. In general, hierarchical color quantization algorithms are faster, whereas partitional ones produce better results provided that they are initialized properly. In this paper, we propose a novel partitional color quantization algorithm based on a binary splitting formulation of MacQueen’s online k-means algorithm. Unlike MacQueen’s original algorithm, the proposed algorithm is both deterministic and free of initialization. Experiments on a diverse set of public test images demonstrate that the proposed algorithm is significantly faster than two popular batch k-means algorithms while yielding nearly identical results. In other words, unlike previously proposed k-means variants, our algorithm addresses both the initialization and acceleration issues of k-means without sacrificing the simplicity of the algorithm. The presented algorithm may be of independent interest as a general-purpose clustering algorithm.

Multi-level Kapur’s thresholding using whale optimization and social group optimization for brain MRI image segmentation

Image thresholding is the vital part for segmentation of a region in image. A fresh technique of multilevel thresholding based on the whale optimizer and social group optimizer are in fact recommended in this particular paper. New Optimization techniques inspired are actually affected as a result of the interpersonal as well as the backpacking behaviour of all of the whales. The process of multi-level thresholding problem is applied through the Kapur’s entropy and metaheuristic algorithms. The suggested strategy is tried on brain MRI test image set. The suggested methods are next in contrast with particle swarm optimization (PSO) and differential evaluation (DE) constructed multilevel thresholding techniques and other optimization techniques in the literature. The segmentation really calculated with evaluation measures like Sensitivity, Specificity, and Dice Score (DS).

An EfficientNet-based modified sigmoid transform for enhancing dermatological macro-images of melanoma and nevi skin lesions

Background and objective: During the initial stages, skin lesions may not have sufficient intensity difference or contrast from the background region on dermatological macro-images. The lack of proper light exposure at the time of capturing the image also reduces the contrast. Low contrast between lesion and background regions adversely impacts segmentation. Enhancement techniques for improving the contrast between lesion and background skin on dermatological macro-images are limited in the literature. An EfficientNet-based modified sigmoid transform for enhancing the contrast on dermatological macro-images is proposed to address this issue.Methods: A modified sigmoid transform is applied in the HSV color space. The crossover point in the modified sigmoid transform that divides the macro-image into lesion and background is predicted using a modified EfficientNet regressor to exclude manual intervention and subjectivity. The Modified EfficientNet regressor is constructed by replacing the classifier layer in the conventional EfficientNet with a regression layer. Transfer learning is employed to reduce the training time and size of the dataset required to train the modified EfficientNet regressor. For training the modified EfficientNet regressor, a set of value components extracted from the HSV color space representation of the macro-images in the training dataset is fed as input. The corresponding set of ideal crossover points at which the values of Dice similarity coefficient (DSC) between the ground-truth images and the segmented output images obtained from Otsu’s thresholding are maximum, is defined as the target.Results: On images enhanced with the proposed framework, the DSC of segmented results obtained by Otsu’s thresholding increased from 0.68 ± 0.34 to 0.81 ± 0.17.Conclusions: The proposed algorithm could consistently improve the contrast between lesion and background on a comprehensive set of test images, justifying its applications in automated analysis of dermatological macro-images.

Deep Reinforcement Learning with Automated Label Extraction from Clinical Reports Accurately Classifies 3D MRI Brain Volumes.

Image classification is probably the most fundamental task in radiology artificial intelligence. To reduce the burden of acquiring and labeling data sets, we employed a two-pronged strategy. We automatically extracted labels from radiology reports in Part 1. In Part 2, we used the labels to train a data-efficient reinforcement learning (RL) classifier. We applied the approach to a small set of patient images and radiology reports from our institution. For Part 1, we trained sentence-BERT (SBERT) on 90 radiology reports. In Part 2, we used the labels from the trained SBERT to train an RL-based classifier. We trained the classifier on a training set of [Formula: see text] images. We tested on a separate collection of [Formula: see text] images. For comparison, we also trained and tested a supervised deep learning (SDL) classification network on the same set of training and testing images using the same labels. Part 1: The trained SBERT model improved from 82 to [Formula: see text] accuracy. Part 2: Using Part 1's computed labels, SDL quickly overfitted the small training set. Whereas SDLshowed the worst possible testing set accuracy of 50%, RL achieved [Formula: see text] testing setaccuracy, with a [Formula: see text]-value of [Formula: see text]. We have shown the proof-of-principle application of automated label extraction from radiological reports. Additionally, we have built on prior work applying RL to classification using these labels, extending from 2D slices to entire 3D image volumes. RL has again demonstrated a remarkable ability to train effectively, in a generalized manner, and based on small training sets.

Comparison of Deep Learning Models for the Classification of Noctilucent Cloud Images

Optically thin layers of tiny ice particles near the summer mesopause, known as noctilucent clouds, are of significant interest within the aeronomy and climate science communities. Ground-based optical cameras mounted at various locations in the arctic regions collect the dataset during favorable summer times. In this paper, first, we compare the performances of various deep learning-based image classifiers against a baseline machine learning model trained with support vector machine (SVM) algorithm to identify an effective and lightweight model for the classification of noctilucent clouds. The SVM classifier is trained with histogram of oriented gradient (HOG) features, and deep learning models such as SqueezeNet, ShuffleNet, MobileNet, and Resnet are fine-tuned based on the dataset. The dataset includes images observed from different locations in northern Europe with varied weather conditions. Second, we investigate the most informative pixels for the classification decision on test images. The pixel-level attributions calculated using the guide back-propagation algorithm are visualized as saliency maps. Our results indicate that the SqueezeNet model achieves an F1 score of 0.95. In addition, SqueezeNet is the lightest model used in our experiments, and the saliency maps obtained for a set of test images correspond better with relevant regions associated with noctilucent clouds.