Jiri Prinosil

Digital identification tags for medical fundus images for tele-ophthalmology applications

This paper proposes a method of inserting a digital pattern having patient identity in the medical image without tampering the medical information of the image. To attain imperceptible insertion of the digital pattern a frequency domain approach is used in the mid frequency band of the discrete cosine transform. The original medical image and the stego-image is compared and analyzed for all features and also tested for retaining of all features and medical information. Blood vessels have been extracted from the original and stego image and it has been established from experimental results that the features remains unaltered. Texture features also has been analyzed and experimental results indicates that the variation in the texture features is minimal and do not affect the medical information. The correlation of the features extracted is above 0.99 indicating the insertion of the digital pattern did not cause any loss of medical information in the image.

Blind face indexing in video

This paper deals with the proposal of blind face indexing system for real videos with no prior knowledge. The system consists of face and facial features detection and tracking stage and face clustering stage. Viola-Jones object detector for face detection, pictorial structures for facial feature detection and Lucas-Kanade optical flow for facial feature tracking are used. Average face models based on popular SIFT features are applied for face clustering. The system operates well on discussion talks or records of videoconferences.

An automatic emotion recognizer using MFCCs and Hidden Markov Models

In this paper, the proficiency of continuous Hidden Markov Models to recognize emotions from speech signals has been investigated. Unlike the existing work which considers prosodic features for automatic emotion recognition, this work proposes the effectiveness of the phonetic features of speech particularly, Mel-Frequency Cepstral Coefficients which improves the accuracy with reduced feature set. The continuous speech emotional utterances used in this work have been taken from the SAVEE emotional corpus. The Hidden Markov Model Toolkit (HTK) version 3.4.1 was utilized for extraction of the acoustic features as well as generation of the models. Optimizing the acoustic and pre-processing parameters along with the number of states and transition probabilities of the Markov Models, the trials give us an average accuracy of 78% and highest accuracy of 91.25% for four emotions sadness, surprise, fear and disgust.

Exudates detection in digital fundus image using edge based method & strategic thresholding

The paper proposes an image processing algorithm to accurately detect the presence of exudates in Fundus images. For Diabetic Retinopathy, presence of exudates in the fundus image marks the beginning of vision loss and hence detecting the exudates accurately and efficiently is of prime concern. The proposed algorithm is a strategic method which removes false detection because of noise generated for different reasons. Using a strategic combination of two independent approaches based on threshold and edge detection helps in eliminating all possible types of noises leading to false exudates that may have crept in. Hence, this method of detecting the exudates has an advantage of increased accuracy. Experimental results indicate that this method has a clear advantage of accuracy in terms of exudates detection in the digital Fundus image without compromising the computational time.

Local descriptors based face recognition engine for video surveillance systems

The paper deals with the local descriptors based face recognition engine for video surveillance systems, where LBP and POEM face descriptors are involved. The FERET database is used for optimal parameters setting with regard to ratio of recognition rate and computational time. The multiple-local-thresholds mechanism is proposed for face template generation and classification task. The engine is verified by using the PubFig database and own database of images from a real surveillance system. The achieved results are suitable to be used in a video surveillance system operating in real-time.

Digital modulation classification based on characteristic features and GentleBoost algorithm

The paper describes algorithm for the classification of digital modulations and its testing with disturbed signals. 2ASK, 2FSK, 4FSK, MSK, BPSK, QPSK, 8PSK and 16QAM were chosen for recognition as the best-known digital modulations used in modern communication technologies. The method designed uses ten features computed from parameters of recognized signal such as instantaneous amplitude, instantaneous phase, instantaneous frequency and spectrum characteristic. The GentleBoost algorithm was used to analyze the features and classify the modulations. We used multipath fading channel to model signal propagation and disturbed the signal by white Gaussian noise for the purpose of testing the algorithm.

Wavelet based robust watermarking scheme for copyright enforcement and integrity control in tele-ophthalmology

Tele-medicine permits medical images to be transmitted among health care centers through the insecure open networks for clinical investigation and for improved healthcare stipulations. There is a growing need for copyright enforcement technologies in these networked multimedia systems. This paper proposes an advanced and modified method to watermark information into a biomedical retinal cover image which aims to provide a robust and adaptive system for medical image protection against authentication and copyright infringement issues in tele-ophthalmological applications. Watermark is embedded in low frequency band using Singular values. Both methods strategically combined increase the level of security of proposed method. In short, a as a new SVD-DWT watermarking algorithm which is robust against various attacks is presented. Performance of the proposed method has been evaluated by assessing the method on metrics MSE (Mean Square Error), PSNR (Peak Signal to Noise Ratio) and NC (Normalized Correlation). Experimental results illustrate the suitability of the algorithm for application in tele-ophthalmology.

An efficient imaging technique for automated macula localization from fundus images

Localization of macula from fundus image plays an important role to design an automated screening tool for detection of retinal diseases. The similar color and texture of red lesions act as a bottleneck in accurate localization of macula in the fundus image. This paper presents a computer vision algorithm for automated and efficient localization of macula from low contrast and diabetic retinopathy affected fundus images. A statistical based model is used to detect macula in a specified region of fundus image which is designed using the geometric features of optic disc. The performance of the proposed algorithm of macula detection was tested on 200 normal/affected fundus images and results are significant. The computational efficiency and accurate localization of macula makes the proposed method competent enough to be used as a part of an automated screening tool for detection of retinal diseases.

Grading of colorectal cancer using histology images

This paper proposed an automated system for grading of colorectal cancer using image processing method. Almost, half a million people die every year due to colon cancer. Histopathological tissue analysis is a common method for its detection, which needs an expert pathologist. Screening for this cancer is effective for prevention as well as early detection. The method proposed segment the glands automatically by using intensity based thresholding and organizational properties for classification. In existing literature, the majority of studies based on gland segmentation in healthy or benign samples, but rarely on intermediate or high grade cancer. Unlike most of the existing methods this system is fully automated and grades the images as benign healthy, benign adenomatous, moderately differentiated malignant and poorly differentiated malignant. The proposed method achieves overall accuracy of 81% when tested on 165 histology images.

An efficient automatic intensity based method for detection of macula in retinal images

This paper demonstrates an automatic proficient intensity based approach for automatic detection and extraction of macula from the retinal fundus images in the field of teleopthalmology. Detection of the macula from a retinal image is an indispensable step for developing automated screening system for ophthalmic pathologies. For eye diagnosis, digital fundus images have become significant, thus opening up the possibility of applying digital image processing techniques in ocular fundus images to facilitate and improve diagnosis. The paper proposes a method for automatic and efficient detection and extraction of macula, based on the reference position of the optical disc from the fundus images. Further, suitable morphological steps are performed on the region of interest obtained with the help of centroid of optical disc to extract the exact location of macula. The proposed algorithm is computationally simple, efficient and can be used to assist ophthalmologists in as a diagnostic tool for screening of various eye related ailments. The experimental results indicate that this method of detection of macula is highly accurate and efficient.