Mahdieh Poostchi

CNN-based image analysis for malaria diagnosis

Malaria is a major global health threat. The standard way of diagnosing malaria is by visually examining blood smears for parasite-infected red blood cells under the microscope by qualified technicians. This method is inefficient and the diagnosis depends on the experience and the knowledge of the person doing the examination. Automatic image recognition technologies based on machine learning have been applied to malaria blood smears for diagnosis before. However, the practical performance has not been sufficient so far. This study proposes a new and robust machine learning model based on a convolutional neural network (CNN) to automatically classify single cells in thin blood smears on standard microscope slides as either infected or uninfected. In a ten-fold cross-validation based on 27,578 single cell images, the average accuracy of our new 16-layer CNN model is 97.37%. A transfer learning model only achieves 91.99% on the same images. The CNN model shows superiority over the transfer learning model in all performance indicators such as sensitivity (96.99% vs 89.00%), specificity (97.75% vs 94.98%), precision (97.73% vs 95.12%), F1 score (97.36% vs 90.24%), and Matthews correlation coefficient (94.75% vs 85.25%).

Feature Selection for Appearance-based Vehicle Tracking in Geospatial Video

Current video tracking systems often employ a rich set of intensity, edge, texture, shape and object level features combined with descriptors for appearance modeling. This approach increases tracker robustness but is compu- tationally expensive for realtime applications and localization accuracy can be adversely affected by including distracting features in the feature fusion or object classification processes. This paper explores offline feature subset selection using a filter-based evaluation approach for video tracking to reduce the dimensionality of the feature space and to discover relevant representative lower dimensional subspaces for online tracking. We com- pare the performance of the exhaustive FOCUS algorithm to the sequential heuristic SFFS, SFS and RELIEF feature selection methods. Experiments show that using offline feature selection reduces computational complex- ity, improves feature fusion and is expected to translate into better online tracking performance. Overall SFFS and SFS perform very well, close to the optimum determined by FOCUS, but RELIEF does not work as well for feature selection in the context of appearance-based object tracking.

Semantic Depth Map Fusion for Moving Vehicle Detection in Aerial Video

Wide area motion imagery from an aerial platform offers a compelling advantage in providing a global picture of traffic flows for transportation and urban planning that is complementary to the information from a network of ground-based sensors and instrumented vehicles. We propose an automatic moving vehicle detection system for wide area aerial video based on semantic fusion of motion information with projected building footprint information to significantly reduce the false alarm rate in urban scenes with many tall structures. Motion detections are obtained using the flux tensor and combined with a scene level depth mask to identify tall structures using height information derived from a dense 3D point cloud estimated using multiview stereo from the same source imagery or a prior model. The trace of the flux tensor provides robust spatio-temporal information of moving edges including the motion of tall structures caused by parallax effects. The parallax induced motions are filtered out by incorporating building depth maps obtained from dense urban 3D point clouds. Using a level-set based geodesic active contours framework, the coarse thresholded tall structures depth masks evolved and stopped at the actual building boundaries. Experiments are carried out on a cropped 2k × 2k region of interest for 200 frames from Albuquerque urban aerial imagery. An average precision of 83% and recall of 76% have been reported using an object-level detection performance evaluation method.

Efficient GPU implementation of the integral histogram

The integral histogram for images is an efficient preprocessing method for speeding up diverse computer vision algorithms including object detection, appearance-based tracking, recognition and segmentation. Our proposed Graphics Processing Unit (GPU) implementation uses parallel prefix sums on row and column histograms in a cross-weave scan with high GPU utilization and communication-aware data transfer between CPU and GPU memories. Two different data structures and communication models were evaluated. A 3-D array to store binned histograms for each pixel and an equivalent linearized 1-D array, each with distinctive data movement patterns. Using the 3-D array with many kernel invocations and low workload per kernel was inefficient, highlighting the necessity for careful mapping of sequential algorithms onto the GPU. The reorganized 1-D array with a single data transfer to the GPU with high GPU utilization, was 60 times faster than the CPU version for a 1K ×1K image reaching 49 fr/sec and 21 times faster for 512×512 images reaching 194 fr/sec. The integral histogram module is applied as part of the likelihood of features tracking (LOFT) system for video object tracking using fusion of multiple cues.

Moving object detection for vehicle tracking in Wide Area Motion Imagery using 4D filtering

Most Wide Area Motion Imagery (WAMI) based trackers use motion based cueing for detecting and tracking moving objects. The results are very high false alarm rates in urban environments with tall structures due to parallax effects. This paper proposes an accurate moving object detection method using a precise orthorectification approach for ground stabilization combined with accurate multiview depth maps to reduce the number of false positives induced by parallax effects by 90 percent. Proposed hybrid moving vehicle detection approach for large scale aerial urban imagery is based on fusion of motion detection mask obtained from median-based background subtraction and tall structures height mask provided by image depth map information. Using buildings mask enables us to improve the object level detection accuracy in terms of F-measure by 57 percent from 22.2% to 79.2%.

Realtime motion detection based on the spatio-temporal median filter using GPU integral histograms

Motion detection using background modeling is a widely used technique in object tracking. To meet the demands of real-time multi-target tracking applications in large and/or high resolution imagery fast parallel algorithms for motion detection are desirable. One common method for background modeling is to use an adaptive 3D median filter that is updated appropriately based on the video sequence. We describe a parallel 3D spatiotemporal median filter algorithm implemented in CUDA for many core Graphics Processing Unit (GPU) architectures using the integral histogram as a building block to support adaptive window sizes. Both 2D and 3D median filters are also widely used in many other computer vision tasks like denoising, segmentation, and recognition. Although fast sequential median algorithms exist, improving performance using parallelization is attractive to reduce the time needed for motion detection in order to support more complex processing in multi-target tracking systems, large high resolution aerial video imagery and 3D volumetric processing. Results show the frame rate of the GPU implementation was 60 times faster than the CPU version for a 1K x 1K image reaching 49 fr/sec and 21 times faster for 512 x 512 frame sizes reaching 194 fr/sec. We characterize performance of the parallel 3D median filter for different image sizes and varying number of histogram bins and show selected results for motion detection.

The Thermal Infrared Visual Object Tracking VOT-TIR2016 Challenge Results

The Thermal Infrared Visual Object Tracking challenge 2015, VOT-TIR2015, aims at comparing short-term single-object visual trackers that work on thermal infrared (TIR) sequences and do not apply pre-learned models of object appearance. VOT-TIR2015 is the first benchmark on short-term tracking in TIR sequences. Results of 24 trackers are presented. For each participating tracker, a short description is provided in the appendix. The VOT-TIR2015 challenge is based on the VOT2013 challenge, but introduces the following novelties: (i) the newly collected LTIR (Link -- ping TIR) dataset is used, (ii) the VOT2013 attributes are adapted to TIR data, (iii) the evaluation is performed using insights gained during VOT2013 and VOT2014 and is similar to VOT2015.

Spatial pyramid context-aware moving vehicle detection and tracking in urban aerial imagery

Persistent detection and tracking of moving vehicles in airborne imagery provide indispensable information for many traffic surveillance applications including traffic monitoring and management, navigation systems, activity recognition and event detection. This paper presents a collaborative Spatial Pyramid Context-aware detection and Tracking system (SPCT) for moving vehicles in dense urban aerial imagery. The proposed system is composed of one master tracker that usually relies on visual object features and two auxiliary trackers based on object temporal motion information that will be called dynamically to assist master tracker. SPCT utilizes image spatial context at different level to make the video tracking system resistant to occlusion, background noise and improve target localization accuracy. We chose a pre-selected seven-channel complementary features including RGB color, intensity and spatial pyramid of HoG (PHoG) and exploit integral histogram as building block to meet the demands of real-time performance. The extensive experiments on ARGUS and ABQ wide aerial video and comparison with state-of-the-art single object trackers confirm that combining complementary tracking cues in an intelligent fusion framework is essential to address the challenges of persistent tracking in low frame rate Wide Aerial Motion Imagery (WAMI).

Detecting and segmenting overlapping red blood cells in microscopic images of thin blood smears.

Automated image analysis of slides of thin blood smears can assist with early diagnosis of many diseases. Automated detection and segmentation of red blood cells (RBCs) are prerequisites for any subsequent quantitative highthroughput screening analysis since the manual characterization of the cells is a time-consuming and error-prone task. Overlapping cell regions introduce considerable challenges to detection and segmentation techniques. We propose a novel algorithm that can successfully detect and segment overlapping cells in microscopic images of stained thin blood smears. The algorithm consists of three steps. In the first step, the input image is binarized to obtain the binary mask of the image. The second step accomplishes a reliable cell center localization that utilizes adaptive meanshift clustering. We employ a novel technique to choose an appropriate bandwidth for the meanshift algorithm. In the third step, the cell segmentation purpose is fulfilled by estimating the boundary of each cell through employing a Gradient Vector Flow (GVF) driven snake algorithm. We compare the experimental results of our methodology with the state-of-the-art and evaluate the performance of the cell segmentation results with those produced manually. The method is systematically tested on a dataset acquired at the Chittagong Medical College Hospital in Bangladesh. The overall evaluation of the proposed cell segmentation method based on a one-to-one cell matching on the aforementioned dataset resulted in 98% precision, 93% recall, and 95% F1-score index.