To Örebro University

Convolutional neural network (CNN)-based methods have achieved state-of-the-art performance in solving several complex computer vision problems including assessment of diabetic retinopathy (DR). Despite this, CNN-based methods are often criticized as “black box” methods for providing limited to no understanding about their internal functioning. In recent years there has been an increased interest to develop explainable deep learning models, and this paper is an effort in that direction in the context of DR. Based on one of the best performing method called Grad-CAM++, we propose Advanced Grad-CAM++ to provide further improvement in visual explanations of CNN model predictions (when compared to Grad-CAM++), in terms of better localization of DR pathology as well as explaining occurrences of multiple DR pathology types in a fundus image. By keeping all the layers and operations as is, the proposed method adds an additional non-learnable bilateral convolutional layer between the input image and the very first learnable convolutional layer of Grad-CAM++. Experiments were conducted on fundus images collected from publicly available sources namely EyePACS and DIARETDB1. Intersection over Union (IoU) score between the ground truth and heatmap produced by the methods were used to quantitatively compare the performance.The overall IoU score for Advanced Grad-CAM++ is 0.179, whereas for Grad-CAM++ it is score 0.161. Thus an 11.18% improvement in agreement with the ground truths by the proposed method is inferable.

Diabetic foot ulcers (DFU) are one of the major health complications for people with diabetes. It may cause limb amputation or lead to life-threatening situations if not detected and treated properly at an early stage. A diabetic patient has a 15–25% chance of developing DFU at a later stage in his or her life if proper foot care is not taken. Because of these high-risk factors, patients with diabetes need to have regular checkups and medications which cause a huge financial burden on both the patients and their families. Hence, the necessity of a cost-effective, re-mote, and fitting DFU diagnosis technique is imminent. This paper presents a convolutional neural network (CNN)-based approach for the automated detection of diabetic foot ulcers from the pictures of a patient’s feet. ResNet50 is used as the backbone of the Faster R-CNN which performed better than the original Faster R-CNN that uses VGG16. A total of 2000 images from the Diabetic Foot Ulcer Grand Challenge 2020 (DFUC2020) dataset have been used for the experiment. The proposed method obtained precision, recall, F1-score, and mean average precision of 77.3%, 89.0%, 82.7%, and 71.3%, respectively, in DFU detection which is better than results obtained by the original Faster R-CNN.

The Green Area Factor(GYF) is an aggregate norm used as an index to quantify how much eco-efficient surface exists in a given area. Although the GYF is a single number, it expresses several different contributions of natural objects to the ecosystem. It is used as a planning tool to create and manage attractive urban environments ensuring the existence of required green/blue elements. Currently, the GYF model is gaining rapid attraction by different communities. However, calculating the GYF value is challenging as significant amount of manual effort is needed. In this study, we present a novel approach for automatic extraction of the GYF value from aerial imagery using semantic segmentation results. For model training and validation a set of RGB images captured by Drone imaging system is used. Each image is annotated into trees, grass, soil/open surface, building, and road. A modified U-net deep learning architecture is used for the segmentation of various objects by classifying each pixel into one of the semantic classes. From the segmented image we calculate the class-wise fractional area coverages that are used as input into the simplified GYF model called Sundbyberg for calculating the GYF value. Experimental results yield that the deep learning method provides about 92% mean IoU for test image segmentation and corresponding GYF value is 0.34.

This paper aims to propose a novel method to detect multiple types of image forgery. The method uses Local Binary Pattern (LBP) as a descriptive feature of the image patches. A uniquely designed convolutional neural network (LBPNet) is proposed where four VGG style blocks are used followed by a support vector machine (SVM) classifier. It uses ‘Swish’ activation function, ‘Adam’ optimizing function, a combination of ‘Binary Cross-Entropy’ and ‘Squared Hinge’ as the loss functions. The proposed method is trained and tested on 111,350 image patches generated from phase-I of IEEE IFS-TC Image Forensics Challenge dataset. Once trained, the results reveal that training such network with computed LBP patches of real and forged image can produce 98.96% validation and 98.84% testing accuracy with area under the curve (AUC) score of 0.988. The experimental result proves the efficacy of the proposed method with respect to the most state-of-the-art techniques.

Image forgery/manipulation is one of the most alarming topics and becomes a major concern about different social media platforms regarding one’s privacy and safety. Therefore, the detection of the manipulated images is of immense interest to the researchers in the recent years. Despite the availability of numerous image forgery detection (IFD) datasets, very few particularly address the actual challenge by collecting the manipulated images from real-world scenario, e.g., collection of images from social media. Consequently, the contextual knowledge behind using the manipulated images remains unachieved. In order to address these issues, we propose an indigenous social media image forgery detection database, naming SMIFD-1000. This dataset provides rich annotations from several aspects: (a) image level: image regions that helps to classify pixel-level information; (b) forgery type: provide rich information about manipulation and (c) target and motif of manipulations: provide contextual rich knowledge about manipulation, which is significantly important from the perspective of social science. Finally, we would examine and benchmark the effectiveness of several publicly available algorithms on this dataset to demonstrate its usefulness. Results show that the dataset is highly challenging and will serve as an important benchmark for the existing and future IFD algorithms. 

Registration of retinal image is a crucial and fundamental step in several medical diagnoses. In this paper we propose an innovative method for retinal image registration. The method applies log-polar transform to approximate the difference in scale and orientation among images. A novel descriptor named Combined Local Haar of Bifurcation points (CLHB) is proposed for robust description and precise matching of retinal bifurcation and cross-over points. Experiments are performed on retinal image registration datasets collected from private and public sources and consisting of a total of 484 fundus photographs (i.e. 242 pairs). The proposed method has been compared with the state-of-the-art Generalized Dual-Bootstrap Iterative Closest Point (GDP ICP), Hernandez-Matas et al., Saha et al., and Chen et al.’s methods and has been found to outperform them with a clear margin. On the publicly available FIRE dataset, our proposed method is found 2% more accurate than the best performing Saha et al.’s method. On the private dataset the method is found to be about 3% more accurate than the best performing method.

Age-related Macular degeneration (AMD) is the third leading cause of incurable acute central vision loss. Optical coherence tomography (OCT) is a diagnostic process used for both AMD and diabetic macular edema (DME) detection. Spectral-domain OCT (SD-OCT), an improvement of traditional OCT, has revolutionized assessing AMD for its high acquiring rate, high efficiency, and resolution. To detect AMD from normal OCT scans many techniques have been adopted. Automatic detection of AMD has become popular recently. The use of a deep Convolutional Neural Network (CNN) has helped its cause vastly. Despite having achieved better performance, CNN models are often criticized for not giving any justification in decision-making. In this paper, we aim to visualize and critically analyze the decision of CNNs in context-based AMD detection. Multiple experiments were done using the DUKE OCT dataset, utilizing transfer learning in Resnet50 and Vgg16 model. After training the model for AMD detection, Gradient-weighted Class Activation Mapping (Grad-Cam) is used for feature visualization. With the feature mapped image, each layer mask was compared. We have found out that the Outer Nuclear layer to the Inner segment myeloid (ONL-ISM) has more predominance about 17.13% for normal and 6.64% for AMD in decision making.

In the field of cultural heritage, applied dyes on textiles are studied to explore their great artistic and historic values. Dye analysis is essential and important to plan correct restoration, preservation and display strategy in museums and art galleries. However, most of the existing diagnostic technologies are destructive to the historical objects. In contrast to that, spectral reflectance imaging is potential as a non-destructive and spatially resolved technique. There have been hardly any studies in classification of dyes in textile fibers using spectral imaging. In this study, we show that spectral imaging with machine learning technique is capable in preliminary screening of dyes into the natural or synthetic class. At first, sparse logistic regression algorithm is applied on reflectance data of dyed fibers to determine some discriminating bands. Then support vector machine algorithm (SVM) is applied for classification considering the reflectance of the selected spectral bands. The results show nine selected bands in short wave infrared region (SWIR, 1000–2500 nm) classify dyes with 97.4% accuracy (kappa 0.94). Interestingly, the results show that fairly accurate dye classification can be achieved using the bands at 1480nm, 1640 nm, and 2330 nm. This indicates possibilities to build an inexpensive handheld screening device for field studies.

Segmentation of retinal blood vessels is a very important diagnostic procedure in ophthalmology. Segmenting blood vessels in the presence of pathological lesions is a majorchallenge. In this paper, an innovative approach to segment the retinal blood vessel in thepresence of pathology is proposed. The method combines both supervised and unsupervised approaches in the retinal imaging context. Two innovative descriptors named localHaar pattern and modified speeded up robust features are also proposed. Experiments areconducted on three publicly available datasets named: DRIVE, STARE and CHASE DB1,and the proposed method has been compared against the state-of-the-art methods. Theproposed method is found about 1% more accurate than the best performing supervisedmethod and 2% more accurate than the state-of-the-art Nguyen et al.’s method.

Segmentation of retinal blood vessels is an important diagnostic procedure in ophthalmology. In this paper we propose an automated blood vessels segmentation method that combines both supervised and un-supervised approaches. A novel descriptor named Local Haar Pattern (LHP) is proposed to describe retinal pixel of interest. The performance of the proposed method has been evaluated on three publicly available DRIVE, STARE and CHASE_DB1 datasets. The proposed method achieves an overall segmentation accuracy of 96%, 96% and 95% respectively on DRIVE, STARE, and CHASE DB1 datasets, which are better than the state-of-the-art methods.