Shruthi Gowda

The (computer) VISIONary!

Defect Inspection using Deep Learning [Oct 2017 – present]
NI VISION, National Instruments (NI) R&D, India
Details: Generalized Defect Inspection caters to diverse Industrial Inspection cases. Deep Learning Experiments conducted with various known networks to building new models. Experiments include Classification algorithms; Segmentation using U-net using Keras, TensorFlow and Intel DAAL (for inference on CPU).
Defect Inspection using Pattern Matching and Geometric Matching [Aug 2016 – June 2017]
NI VISION, National Instruments (NI) R&D, India
Details: Defect Inspection is a widely used feature in Industrial Inspection algorithms. Defect maps aid in detecting defects in images and patterns when a template of the same is known. The Matching algorithms provides an overall score for a match (or matches). A more localized scoring mechanism has been developed which provides more information about how each pixel of the match differs from the template and hence is a great tool in detecting defects. The defective part is found from the image using Pattern Matching (feature: pixel intensity) or Geometric matching (feature: edges, curves). The template is now compared with the matched defect image using different metrics to obtain a Defect Map. A defect map is a float-point image that has a score for every pixel. The higher the score, the higher the probability of the pixel being a defect. Additionally a weight map is calculated from the templates and is used to specify weights to suppress noise and false defects in the defect map.
Pattern Matching [Aug 2015 – Jan 2016]
NI VISION, National Instruments (NI) R&D, India
Details: Pattern Matching is a very important tool used in varying applications such as Alignment, Gauging and Inspection. A novel Pyramidal matching algorithm is implemented. The pattern matching process consists of two stages: learning and matching. During the learning stage, the algorithm extracts gray value and/or edge gradient information from the template image. During the matching stage, the pattern matching algorithm extracts gray value and/or edge gradient information from the inspection image. Then, the algorithm finds matches by locating regions in the inspection image where the highest (normalized)cross-correlation is observed. In pyramidal matching, both the image and the template are sampled to smaller spatial resolutions using Gaussian pyramids. The locations are calculated up to sub-pixel accuracy.
Feature Detection and Matching [Aug 2014 – June 2015]
NI VISION, National Instruments (NI) R&D, India
Details: The feature detection has an improved Harris Corner Detector. It follows a pyramidal approach and the corner locations get continuously refined locally while ascending to the high resolution(or original) image. This approach makes the detection quite fast and robust. Feature Descriptors are provided to attain correspondence between images. This feature also provides the homography information and the estimated pose between the views.
1D Barcode Localization and Decoding and Barcode Grading [Sep 2013 – Aug 2014]
NI VISION, National Instruments (NI) R&D, India
Details: The Barcode Reader product which was released on Aug 2014 has Barcode localizing, reading and grading features. The barcodes are located in the image using segmentation and classification. The decoder logic is different for each of then ten different barcode types supported. The grading tests the quality of the barcode given by the standards. This feature is robust and can locate and read multiple barcodes in low contrast images.
Object Tracking Feature [Jan 2013 – Aug 2013]
NI VISION, National Instruments (NI) R&D, India
Details: The Tracker feature which was released on Aug 2013 tracks an object through frames and is invariant to Scale, Shape and rotation of the object. The Algorithm used is a Continuously Adaptive Mean Shift method.
MASTER THESIS : Virtual Viewpoint Reconstruction by View Interpolation in a Multi Camera Network [Jan 2012 – Sep 2012]
Ecole Polytechnique de Louvain School of Engineering (EPL)
UCL, Belgium
Professor: Christophe de Vleeschouwer

Details:The Thesis proposes a method towards reconstruction of dynamic regions (players) of a basketball match in a virtual viewpoint using the images from multiple cameras. A novel synthesis method, which is a mixture of a weak image model based approach and the transfer based approach is used for view synthesis. The player is modelled in 3D with a very simple model, namely a plane; and the color information is used to synthesize the texture of these 3D models. The transfer based approach, which used projective geometry between neighbouring views, is used for estimating player positions. The efforts for camera calibration are reduced as cameras do not need to be strongly calibrated for projective geometry; hence, the proposed method can be easily applied to dynamic events in a large space.
Automated Electrocardiogram Prediction using regression models and Neural Networks [Sep 2011-Dec 2011]
Institute of Information and Communication Technologies, Electronics and Applied Mathematics (ICTEAM)
UCL, Belgium
Details: The main purpose of this project was to implement my Machine Learning knowledge. The project develops several regression models, namely Linear Regression model, kNN Regression model and Radial Basis Function Network (RBFN), to predict one beat of an ECG signal from a given one wave of the signal. The trained regression models are used to predict the next value. The overall performances of the models are then evaluated. Exhaustive methods were implemented (Bootstrap, cross validation, etc..) to find the optimal meta-parameters in both kNN regression and RBFN models.
Medical Imaging : MR Reconstruction, Registration and Segmentation [Sep 2011-Dec 2011]
UCL, Belgium
Details: The purpose of this challenge is to segment a brain lesion using two MR images. First, the T1 image from the signal coming from the MRI equipment reconstructed, then it is registered (Rigid and Non rigid Registration) with a Proton-Density (PD) image of the same patient. The lesion is then segmented (Region growing) on both images and the intersection is computed in order to get a better contour of the lesion.
Feature Extraction Methods for Speech Recognition [Feb 2011-Apr 2011]
Center for Language and Speech Technologies and Applications (TALP)
Details: For recognition, the excitation source and vocal tract filter should be properly separated since only envelope, but not pitch information, is necessary to accomplish recognition task. In order to fulfill this criteria one not only needs to obtain the parameters to clearly identify envelope evolutions but also be able to compact these parameters to a reduced set so that even few parameters will represent the envelope of the spectrum of a given frame. First the input speech waveform is transformed into a sequence of acoustic feature vectors, each vector representing the information in a small time window of the signal. The speech features are then extracted using a new parameterization called MFCC (Mel Frequency Cepstral Coefficients) and is compared with LPC Vocoder (Linear Predictive Coefficients) method.
Fixed point IF Filter implementation in the Radar Receivers for Real Time Signal Processing [Jan 2010-June 2010]
LRDE (Electronics and Radar Development Establishment)
DRDO (Defence Research and Development Organization), India.