Lasse Lensu

The DIARETDB1 diabetic retinopathy database and evaluation protocol

Automatic diagnosis of diabetic retinopathy from digital fundus images has been an active research topic in the medical image processing community. The research interest is justified by the excellent potential for new products in the medical industry and significant reductions in health care costs. However, the maturity of proposed algorithms cannot be judged due to the lack of commonly accepted and representative image database with a verified ground truth and strict evaluation protocol. In this study, an evaluation methodology is proposed and an image database with ground truth is described. The database is publicly available for benchmarking diagnosis algorithms. With the proposed database and protocol, it is possible to compare different algorithms, and correspondingly, analyse their maturity for technology transfer from the research laboratories to the medical practice.

Detection of irregularities in regular patterns

Regular patterns, as defined in this study, are found in areas of industry and science, for example, halftone raster patterns used in the printing industry and crystal lattice structures in solid state physics. The need for quality inspection of products containing regular patterns has aroused interest in the application of machine vision for automatic inspection. Quality inspection typically corresponds to detecting abnormalities, defined as irregularities in this case. In this study, the problem of irregularity detection is described in analytical form and three different detection methods are proposed. All the methods are based on characteristics of the Fourier transform to compactly represent regular information. The Fourier transform enables the separation of regular and irregular parts of an input image. The three methods presented are shown to differ in their generality and computational complexities.

Projector calibration by inverse camera calibration

The accuracy of 3-D reconstructions depends substantially on the accuracy of active vision system calibration. In this work, the problem of video projector calibration is solved by inverting the standard camera calibration work flow. The calibration procedure requires a single camera, which does not need to be calibrated and which is used as the sensor whether projected dots and calibration pattern landmarks, such as the checkerboard corners, coincide. The method iteratively adjusts the projected dots to coincide with the landmarks and the final coordinates are used as inputs to a camera calibration method. The otherwise slow iterative adjustment is accelerated by estimating a plane homography between the detected landmarks and the projected dots, which makes the calibration method fast.

Mottling assessment of solid printed areas and its correlation to perceived uniformity

Mottling is one of the most important printing defects in modern offset printing using coated papers. Mottling can be defined as undesired unevenness in perceived print density. In our research, we have implemented three methods to evaluate print mottle: the standard method, the cluster-based method, and the bandpass method. Our goal was to study the methods presented in literature, and modify them by taking relevant characteristics of the human visual system into account. For comparisons, we used a test set of 20 grey mottle samples which were assessed by both humans and the modified methods. The results show that when assessing low-contrast unevenness of print, humans have diverse opinions about quality, and none of the methods accurately capture the characteristics of human vision.

Constructing Benchmark Databases and Protocols for Medical Image Analysis: Diabetic Retinopathy

We address the performance evaluation practices for developing medical image analysis methods, in particular, how to establish and share databases of medical images with verified ground truth and solid evaluation protocols. Such databases support the development of better algorithms, execution of profound method comparisons, and, consequently, technology transfer from research laboratories to clinical practice. For this purpose, we propose a framework consisting of reusable methods and tools for the laborious task of constructing a benchmark database. We provide a software tool for medical image annotation helping to collect class label, spatial span, and experts confidence on lesions and a method to appropriately combine the manual segmentations from multiple experts. The tool and all necessary functionality for method evaluation are provided as public software packages. As a case study, we utilized the framework and tools to establish the DiaRetDB1 V2.1 database for benchmarking diabetic retinopathy detection algorithms. The database contains a set of retinal images, ground truth based on information from multiple experts, and a baseline algorithm for the detection of retinopathy lesions.

Making Visual Object Categorization More Challenging: Randomized Caltech-101 Data Set

Visual object categorization is one of the most active research topics in computer vision, and Caltech-101 data set is one of the standard benchmarks for evaluating the method performance. Despite of its wide use, the data set has certain weaknesses: i) the objects are practically in a standard pose and scale in the middle of the images and ii) background varies too little in certain categories making it more discriminative than the foreground objects. In this work, we demonstrate how these weaknesses bias the evaluation results in an undesired manner. In addition, we reduce the bias effect by replacing the backgrounds with random landscape images from Google and by applying random Euclidean transformations to the foreground objects. We demonstrate how the proposed randomization process makes visual object categorization more challenging improving the relative results of methods which categorize objects by their visual appearance and are invariant to pose changes. The new data set is made publicly available for other researchers.

Bayesian network model of overall print quality: Construction and structural optimisation

Prediction of overall visual quality based on instrumental measurements is a challenging task. Despite the several proposed models and methods, there exists a gap between the instrumental measurements of print and human visual assessment of natural images. In this work, a computational model for representing and quantifying the overall visual quality of prints is proposed. The computed overall quality should correspond to the human visual quality perception when viewing the printed images. The proposed model is a Bayesian network which connects the objective instrumental measurements to the subjective opinion distribution of human observers. This relationship can be used to score printed images, and additionally, to computationally study the connections of the attributes. A novel graphical learning approach using an iterative evolve-estimate-simulate loop learning the quality model based on psychometric data and instrumental measurements is suggested. The network structure is optimised by applying evolutionary computation (evolve). The estimation of the Bayesian network parameters is within the evolutionary loop. In this loop, the maximum likelihood approach is used (estimate). The stochastic learning process is guided by priors devised from the psychometric subjective experiments (performance through simulation). The model reveals and represents the explanatory factors between its elements providing insight to the psychophysical phenomenon of how observers perceive visual quality and which measurable entities affect the quality perception. By using true data, the design choices are demonstrated. It is also shown that the best-performing network establishes a clear and intuitively correct structure between the objective measurements and psychometric data.

Fusion of Multiple Expert Annotations and Overall Score Selection for Medical Image Diagnosis

Two problems especially important for supervised learning and classification in medical image processing are addressed in this study: i) how to fuse medical annotations collected from several medical experts and ii) how to form an image-wise overall score for accurate and reliable automatic diagnosis. Both of the problems are addressed by applying the same receiver operating characteristic (ROC) framework which is made to correspond to the medical practise. The first problem arises from the typical need to collect the medical ground truth from several experts to understand the underlying phenomenon and to increase robustness. However, it is currently unclear how these expert opinions (annotations) should be combined for classification methods. The second problem is due to the ultimate goal of any automatic diagnosis, a patient-based (image-wise) diagnosis, which consequently must be the ultimate evaluation criterion before transferring any methods into practise. Various image processing methods provide several, e.g., spatially distinct, results, which should be combined into a single image-wise score value. We discuss and investigate these two problems in detail, propose good strategies and report experimental results on a diabetic retinopathy database verifying our findings.

Automated mottling assessment of colored printed areas

Mottling is one of the most significant defects in modern offset printing using coated papers. Mottling can be defined as undesired unevenness in perceived print density. Previous research in the field considered only gray scale prints. In our work, we extend current methodology to color prints. Our goal was to study the characteristics of the human visual system, perform psychometric experiments and develop methods which can be used at industrial level applications. We developed a method for color prints and extensively tested it with a number of experts and laymen. Suggested approach based on pattern-color perception separability proved to correlate with the human evaluation well.

Unsupervised object discovery via self-organisation

Object discovery in visual object categorisation (VOC) is the problem of automatically assigning class labels to objects appearing in given images. To achieve state-of-the-art results in this task, a large set of positive and negative training images from publicly available benchmark data sets have been used to train discriminative classification methods. The immediate drawback of these methods is the requirement of a vast amount of labelled data. Therefore, the ultimate challenge for visual object categorisation has been recently exposed: unsupervised object discovery, also called unsupervised VOC (UVOC), where the selection of the number of classes and the assignments of given images to these classes are performed automatically. The problem is very challenging and hitherto only a few methods have been proposed. These methods are based on the popular bag-of-features approach and clustering to automatically form the classes. In this paper, we adopt the self-organising principle and replace clustering with the self-organising map (SOM) algorithm. Our method provides results comparable to the state of the art and its advantages, such as non-sensitivity against codebook histogram normalisation, advocate its usage in unsupervised object discovery.