Maryam Rahnemoonfar

Deep Count: Fruit Counting Based on Deep Simulated Learning

Recent years have witnessed significant advancement in computer vision research based on deep learning. Success of these tasks largely depends on the availability of a large amount of training samples. Labeling the training samples is an expensive process. In this paper, we present a simulated deep convolutional neural network for yield estimation. Knowing the exact number of fruits, flowers, and trees helps farmers to make better decisions on cultivation practices, plant disease prevention, and the size of harvest labor force. The current practice of yield estimation based on the manual counting of fruits or flowers by workers is a very time consuming and expensive process and it is not practical for big fields. Automatic yield estimation based on robotic agriculture provides a viable solution in this regard. Our network is trained entirely on synthetic data and tested on real data. To capture features on multiple scales, we used a modified version of the Inception-ResNet architecture. Our algorithm counts efficiently even if fruits are under shadow, occluded by foliage, branches, or if there is some degree of overlap amongst fruits. Experimental results show a 91% average test accuracy on real images and 93% on synthetic images.

Restoration of Arbitrarily Warped Historical Document Images Using Flow Lines

Historical documents frequently suffer from arbitrary geometric distortions (warping and folds) due to storage conditions, use and to, some extent, the printing process of the time. In addition, page curl can be prominent due to the scanning technique used. Such distortions adversely affect OCR and print-on-demand quality. Previous approaches to geometric restoration either focus only on the correction of page curl or require supplementary information obtained by additional scanning hardware -- not practical for existing scans. This paper presents a new approach to detect and restore arbitrary warping and folds, in addition to page curl. Warped text lines and the smooth deformation between them are precisely modelled as primary and secondary flow lines that are then restored to their original linear shape. Preliminary, but representative, experimental results, in comparison to a leading page curl removal method and an industry-standard commercial system, demonstrate the effectiveness of the proposed method.

Automatic polar ice thickness estimation from SAR imagery

Global warming has caused serious damage to our environment in recent years. Accelerated loss of ice from Greenland and Antarctica has been observed in recent decades. The melting of polar ice sheets and mountain glaciers has a considerable influence on sea level rise and altering ocean currents, potentially leading to the flooding of the coastal regions and putting millions of people around the world at risk. Synthetic aperture radar (SAR) systems are able to provide relevant information about subsurface structure of polar ice sheets. Manual layer identification is prohibitively tedious and expensive and is not practical for regular, longterm ice-sheet monitoring. Automatic layer finding in noisy radar images is quite challenging due to huge amount of noise, limited resolution and variations in ice layers and bedrock. Here we propose an approach which automatically detects ice surface and bedrock boundaries using distance regularized level set evolution. In this approach the complex topology of ice and bedrock boundary layers can be detected simultaneously by evolving an initial curve in radar imagery. Using a distance regularized term, the regularity of the level set function is intrinsically maintained that solves the reinitialization issues arising from conventional level set approaches. The results are evaluated on a large dataset of airborne radar imagery collected during IceBridge mission over Antarctica and Greenland and show promising results in respect to hand-labeled ground truth.

Automatic Seagrass Disturbance Pattern Identification on Sonar Images

Natural and anthropogenic disturbances are causing degradation and loss of seagrass cover, often in the form of bare patches (potholes) and propeller-scaring from vessels. Degradation of seagrass habitat has increased significantly in recent years with losses totaling some 110 km2 per year. With seagrass habitat disappearing at historically unprecedented rates, development of new tools for mapping these disturbances is critical to understanding habitat distribution and seagrass abundance. Current methods for mapping seagrass coverage rely on appropriate meteorological conditions (satellite imagery), are high in cost (aerial photography), or lack resolution (in situ point surveys). All of these methods require low turbidity, and none is capable of automatically detecting bare patches (potholes) in seagrass habitat. Sonar-based methods for mapping seagrass can function in high turbidity, and are not affected by meteorological conditions. Here, we present an automatic method for detecting and quantifying potholes in sidescan sonar images collected in a very shallow, highly disturbed seagrass bed. Acoustic studies of shallow seagrass beds (<2 m) are scarce due to traditional approaches being limited by reduced horizontal swath in these depth ranges. The main challenges associated with these sidescan sonar images are random ambient noise and uneven backscatter intensity across the image. Our method combines adaptive histogram equalization and top-hat mathematical morphology transformation to remove image noises and irregularities. Then, boundaries of potholes are detected using optimum binarization as well as closing and erosion mathematical morphology filters. This method was applied to several sonar images taken from the Lower Laguna Madre in Texas at less than 2-m depth. Experimental results in comparison with ground-truthing demonstrated the effectiveness method by identifying potholes with 97% accuracy.

Real-time yield estimation based on deep learning

Crop yield estimation is an important task in product management and marketing. Accurate yield prediction helps farmers to make better decision on cultivation practices, plant disease prevention, and the size of harvest labor force. The current practice of yield estimation based on the manual counting of fruits is very time consuming and expensive process and it is not practical for big fields. Robotic systems including Unmanned Aerial Vehicles (UAV) and Unmanned Ground Vehicles (UGV), provide an efficient, cost-effective, flexible, and scalable solution for product management and yield prediction. Recently huge data has been gathered from agricultural field, however efficient analysis of those data is still a challenging task. Computer vision approaches currently face diffident challenges in automatic counting of fruits or flowers including occlusion caused by leaves, branches or other fruits, variance in natural illumination, and scale. In this paper a novel deep convolutional network algorithm was developed to facilitate the accurate yield prediction and automatic counting of fruits and vegetables on the images. Our method is robust to occlusion, shadow, uneven illumination and scale. Experimental results in comparison to the state-of-the art show the effectiveness of our algorithm.

Two dimensional phase unwrapping of interferometric SAR data by means of wavelet technique

One of the reliable methods in least-squares method is multigrid technique which overcomes the problem of slow convergence and less-accurate of Gauss-Seidel by transforming problem to coarser grid. It makes a pyramid of grids. Each grid has half the resolution of its predecessor. It uses two restriction and prolongation operators called fine-to-coarse and coarse-to-fine operators respectively. In this research, discrete wavelet decomposition and its reconstruction have been applied on the two operators. One of the assumptions made on this operator is that as long as the wavelet transformation decomposes the 2-D signal to one low frequency and three high frequency components, it should converge faster and more accurate than the multigrid method. This is due to the fact that the transformation of only low frequency component would suffice rather than transforming the whole grid to coarser grid. The idea has been implemented and tested on simulation data and the results confirm the assumption. In this paper the results of implementation of various wavelet filters and also multigrid techniques on various simulation data (with and without noise) are presented. In all cases, wavelet techniques have shown improved results than multigrid techniques.

The First Automatic Method for Mapping the Pothole in Seagrass

There is a vital need to map seagrass ecosystems in order to determine worldwide abundance and distribution. Currently there is no established method for mapping the pothole or scars in seagrass. Detection of seagrass with optical remote sensing is challenged by the fact that light is attenuated as it passes through the water column and reflects back from the benthos. Optical remote sensing of seagrass is only possible if the water is shallow and relatively clear. In reality, coastal waters are commonly turbid, and seagrasses can grow under 10 meters of water or even deeper. One of the most precise sensors to map the seagrass disturbance is side scan sonar. Underwater acoustics mapping produces a high definition, two-dimensional sonar image of seagrass ecosystems. This paper proposes a methodology which detects seagrass potholes in sonar images. Side scan sonar images usually contain speckle noise and uneven illumination across the image. Moreover, disturbance presents complex patterns where most segmentation techniques will fail. In this paper, the quality of image is improved in the first stage using adaptive thresholding and wavelet denoising techniques. In the next step, a novel level set technique is applied to identify the pothole patterns. Our method is robust to noise and uneven illumination. Moreover it can detect the complex pothole patterns. We tested our proposed approach on a collection of underwater sonar images taken from Laguna Madre in Texas. Experimental results in comparison with the ground-truth show the efficiency of the proposed method.

Automatic detection and counting of cattle in UAV imagery based on machine vision technology (Conference Presentation)

Beef production is the main agricultural industry in Texas, and livestock are managed in pasture and rangeland which are usually huge in size, and are not easily accessible by vehicles. The current research method for livestock location identification and counting is visual observation which is very time consuming and costly. For animals on large tracts of land, manned aircraft may be necessary to count animals which is noisy and disturbs the animals, and may introduce a source of error in counts. Such manual approaches are expensive, slow and labor intensive. In this paper we study the combination of small unmanned aerial vehicle (sUAV) and machine vision technology as a valuable solution to manual animal surveying. A fixed-wing UAV fitted with GPS and digital RGB camera for photogrammetry was flown at the Welder Wildlife Foundation in Sinton, TX. Over 600 acres were flown with four UAS flights and individual photographs used to develop orthomosaic imagery. To detect animals in UAV imagery, a fully automatic technique was developed based on spatial and spectral characteristics of objects. This automatic technique can even detect small animals that are partially occluded by bushes. Experimental results in comparison to ground-truth show the effectiveness of our algorithm.

Improved Locally Linear Embedding for Big-data Classification

A hyperspectral image provides a multidimensional data consisting of hundreds of spectral dimensions. Even though having an abundance of spectral might seem favorable, classification of hyperspectral data tends to collide with the curse of dimensionality. Therefore, reducing the number of dimensions before classification is always favorable. For this research, the feature extraction method will consist of a nonlinear manifold learning technique named locally linear embedding (LLE). Additionally, another problem that we attempt to overcome is the high computational time required to run manifold learning methods. In order to help overcome this problem, this research compares one implementation of LLE against an improved version that runs much quicker than the original version.

Automatic Ice Surface and Bottom Boundaries Estimation in Radar Imagery Based on Level-Set Approach

Accelerated loss of ice from Greenland and Antarctica has been observed in recent decades. The melting of polar ice sheets and mountain glaciers has considerable influence on sea level rise in a changing climate. Ice thickness is a key factor in making predictions about the future of massive ice reservoirs. The ice thickness can be estimated by calculating the exact location of the ice surface and subglacial topography beneath the ice in radar imagery. Identifying the locations of ice surface and bottom is typically performed manually, which is a very time-consuming procedure. Here, we propose an approach, which automatically detects ice surface and bottom boundaries using distance-regularized level-set evolution. In this approach, the complex topology of ice surface and bottom boundary layers can be detected simultaneously by evolving an initial curve in the radar imagery. Using a distance-regularized term, the regularity of the level-set function is intrinsically maintained, which solves the reinitialization issues arising from conventional level-set approaches. The results are evaluated on a large data set of airborne radar imagery collected during a NASA IceBridge mission over Antarctica and show promising results with respect to manually picked data.