Understanding of Emotional response to image/video attributes(ex. content, quality) is useful in many applications such as perception based video coding, quality assessment, maximizing user experience, etc. During the internship, we concentrated on emotional response to the content of images/video. EEG signals recorded higher ERP(beta frequency range) on brain regions FC5 and FC6 for emotionally troubling images and action-scenes of video. These findings can be used to quantify audience's perception for advertisement/movie.
USB2.0 and OTG testcases were migrated from TLM based environment to new LVP based environment. IP verification was performed. Issues related to device initialization like race condition, memory storage for sending/receiving datapackets, register overwriting, etc were resolved. New APIs for easy generation new testcases was developed and distributed. Regression environment was set and tested.
Computer aided diagnosis(CAD) is useful in rapid medical image analysis. Understanding of how doctor perceive images can help to develop human-like CAD. We tried to understand the gaze pattern of 57 retinal experts with various level of training. Eye-tracking experiments were carried out in 5 different hospitals. Accuracy, response-time, dwell-time, latency, etc were analysed which can be used to learn best gaze pattern to be followed by CAD.
We developed a bottom-up saliency model for DR images which was further used to develop Interactive Selective Enhancement (ISE) of DR lesions. Proposed saliency model was biologically inspired and based on spatially-varying morphological operation. Domain of morphological operation was allowed to vary based on gaussian field observed in human visual system. At 30% saliency our method achieved same f-score as human. Saliency based ISE can be used by retinal experts as a CAD tool.
We developed a Convolutional Neural Network (CNN) based top-down saliency for DR lesion. Architecture of CNN was derived from Itti-Koch saliency model. Half of the weights were initialised with standard orientation, centre-surround filters and others with random filters. CNN fine-tunes bottom-up filters and simultaneously learns new filters. A novel loss function was designed to handle range mismatch problem. AUC achieved for bright and dark lesion was 0.95 and 0.91 respectively. Computed saliency was used to develop Assistive Lesion Emphasis System (ALES).
This is a CNN+RNN architecture trained in a manner where more informative samples are seen for more times. post processing is done using active contour based method. More details coming soon...!!!
OCT images were first converted to frequency domain. Magnitude spectrum was smeared rotationally around n pivot points to generate n maps. These maps were averaged to generate an interference map(IM). IM was multiplied with magnitude spectrum by keeping phase spectrum unaltered. Symmetry was enforced using hilbert transform to ensure real valued reconstruction in spatial domain. Image was converted back to spatial domain.
This was an extension to the in-house work. Machine learning based retinal lesion detection algorithms require huge amount of labelling for training. In the situation where data is unlimited but ground-truth marking is limited, various semi-supervised techniques were devised and tested for learning.
OCT image was first denoised using anisotropic diffusion. Next, image was decomposed into 6 directional bands for 2 scales using steerable pyramid. Band responses of top three bands with highest energy were fused using pixel-wise max operation. Fused responses at 2 scales were multiplied to get clean edges. Vessels were detected by thresholding the summation of all rows and masked out. A polynomial was fitted on the detected edges to get smooth layer. Achieved average error rate was ~1 pixel.
ACM based OD segmentation is challenging task due to presence of vessels. External gradient field is not normal to the OD boundary near vessel region, which makes the convergence of contour difficult. At each pixel an ideal external field was computed and actual field more than 30 degree away from ideal field was masked. This field was converted into quasi-polar domain and spline fitting was performed to predict new field. ACM works well with newly generated gradient field.
Movie was rated on scale of 5 from the movie review dataset of Rotten Tomato using nature language processing techniques. TF-IDF, skip-words, stemming, classification and other techniques were applied. It was a python implementation and achieved ~65% accuracy.
The feature integration theory by Treisman was verified for all 9 testcases presented originally. Four females and four males took part with each one performing 200 trials per testcase.
Images with coloring dots on black background were generated. Neural network has to learn to track the locations of dots in particular color sequence. Images contain 5 to 7 dots of different colors, which models varied sequence-length. A model is combination of CNN and RNN. Image passes through CNN and computed features were fed to RNN. RNN computes location of dot with next color in the sequence. After all dots are tracked a special EOS(end of sequence) token is generated. CNN+RNN model was trained in end-to-end fashion.
The problem is to learn to draw a perfect square or triangle depending on the first input coordinate(polar). For example, if input is (r,a) then draw a square with sides proportional r and if input is (r,-a) then draw a triangle with sides proportional to r. This is done using RNN. This problem can be extended to learn the gaze pattern of the individual depending on the first location (s)he prefers to land in the image.
Tic-tac-toe is the simplest gaming problem where two opponents try to align their own marks(X or O). The problem was defined for mxn board. Policy gradient algorithm in conjunction with 2 layer multilayer perceptron(MLP) was used. After looking at the current arrangement of the board, MLP generates next location for the move and a random agent plays the other part. Reward is seen at the end of the game and backpropogation through time with weighted gradients was used for training.