The emergence of deep-learning methods in different computer vision tasks has proved to offer increased detection, recognition or segmentation accuracy when large annotated image datasets are available. In the case of medical image processing and computer-aided diagnosis within ultrasound images, where the amount of available annotated data is smaller, a natural question arises: are deep-learning methods better than conventional machine-learning methods? How do the conventional machine-learning methods behave in comparison with deep-learning methods on the same dataset? Based on the study of various deep-learning architectures, a lightweight multi-resolution Convolutional Neural Network (CNN) architecture is proposed. It is suitable for differentiating, within ultrasound images, between the Hepatocellular Carcinoma (HCC), respectively the cirrhotic parenchyma (PAR) on which HCC had evolved. The proposed deep-learning model is compared with other CNN architectures that have been adapted by transfer learning for the ultrasound binary classification task, but also with conventional machine-learning (ML) solutions trained on textural features. The achieved results show that the deep-learning approach overcomes classical machine-learning solutions, by providing a higher classification performance.

The early recognition and understanding of the actions performed by pedestrians in traffic scenes leads to an anticipation of pedestrian intentions in advance and helps in the process of collision warning and avoidance in the context of autonomous vehicles. An environment with low visibility conditions such as night-time, fog, heavy rain or smoke increases the number of difficult situations in traffic. A complete and original model for assessing if a pedestrian is engaged in a street cross action using only infrared monocular scene perception is proposed in this paper. The assessment of a street cross action is done by the time series analysis of features like: pedestrian motion, position of pedestrians with respect to the drivable area and their distance with respect to the ego-vehicle. The extraction of these features emerges from the combination of a deep learning based pedestrian detector with an original tracking algorithm, a semantic segmentation of the road surface and a time series long-short term memory network based action recognition. In order to validate the proposed method we introduce a new dataset named CROSSIR. It is formed of pedestrian annotations, action annotations and semantic labels for the road. The CROSSIR dataset is suitable for several common computer vision algorithms: (1) pedestrian detection and tracking algorithms because each pedestrian has a unique identifier over the frames in which it appears; (2) pedestrian action recognition; (3) semantic segmentation of the road pixels in the infrared image.

Hepatocellular Carcinoma (HCC) is the most frequent malignant liver tumor and the third cause of cancer-related deaths worldwide. For many years, the golden standard for HCC diagnosis has been the needle biopsy, which is invasive and carries risks. Computerized methods are due to achieve a noninvasive, accurate HCC detection process based on medical images. We developed image analysis and recognition methods to perform automatic and computer-aided diagnosis of HCC. Conventional approaches that combined advanced texture analysis, mainly based on Generalized Co-occurrence Matrices (GCM) with traditional classifiers, as well as deep learning approaches based on Convolutional Neural Networks (CNN) and Stacked Denoising Autoencoders (SAE), were involved in our research. The best accuracy of 91% was achieved for B-mode ultrasound images through CNN by our research group. In this work, we combined the classical approaches with CNN techniques, within B-mode ultrasound images. The combination was performed at the classifier level. The CNN features obtained at the output of various convolution layers were combined with powerful textural features, then supervised classifiers were employed. The experiments were conducted on two datasets, acquired with different ultrasound machines. The best performance, above 98%, overpassed our previous results, as well as representative state-of-the-art results.

We propose a method for detecting pedestrians in infrared images. The method combines a fast region of interest generator with fast feature pyramid object detection. Knowing the appearance model of pedestrians in infrared images we infer some edge and intensity based filters that generate the regions in which pedestrian hypotheses may appear. On those regions we apply the Aggregated Channel Features introduced by [1]. We train and test the proposed solution on an infrared pedestrian data set and the results show a good detection accuracy and small execution time of about 30fps.

Recent work in monocular pedestrian detection is trying to improve the execution time while keeping the accuracy as high as possible. A popular and successful approach for monocular intensity pedestrian detection is based on the approximation (instead of computation) of image features for multiple scales based on the features computed on set of predefined scales. We port this idea to the infrared domain. Our contributions reside in the combination of four channel features, namely infrared, histogram of gradient orientations, normalized gradient magnitude and local binary patterns with the objective of detecting pedestrians for night vision applications dealing with far infrared sensors. Multiple scale feature computation is done by feature approximation. Another contribution is the study of different formulations for Local Binary Patterns like uniform patterns and rotation invariant patterns and their effect on detection performance. The detection speed is also boosted by the aid of a fast morphological based region of interest generator. We vary the number of approximated scales per octave and study the impact on execution time and accuracy. A reasonable result hits a speed of 18fps with a log average miss rate of 39%.

Object recognition from images is one of the essential problems in automatic image processing. In this paper we focus specifically on nearest neighbor methods, which are widely used in many practical applications, not necessarily related to image data. It has recently come to attention that high dimensional data also exhibit high hubness, which essentially means that some very influential data points appear and these points are referred to as hubs. Unsurprisingly, hubs play a very important role in the nearest neighbor classification. We examine the hubness of various image data sets, under several different feature representations. We also show that it is possible to exploit the observed hubness and improve the recognition accuracy.