Please enable JavaScript.
Coggle requires JavaScript to display documents.
Computational Vision (Human Vision (Photocells - Single cells where a part…
Computational Vision
Human Vision
-
Evolution of eyes - From single cell to multi cells and then curved muti cells. Allows light to be picked up from a bigger area and have direction.
Pinhole camera - All light is focused to a single point for better resolution/brightness. Image is flipped so needs to be flipped again
Rods (120M) - Extremely sensitive to photons (pick up dim light), poor spacial resolution. Many rods to a single Ganglion cell.
Ion Channel - Allows intake of chemicals to the photocell. Normally open. Closes when exposed to light, starts to create a charge unbalance.
Cones (6M) - Only active at high light levels, very precise. One to one with Ganglion. Give colour
Receptive Field - Area where light falls to be detected. More photo-receptors in the centre. Squinting sharpens image to centre.
Ganglion cell - Processes information from photo-receptors. On centre fires when light is in centre. Off centre when off. Both fire weakly if both sections are detected. Allows to see contrast
Visual Pathway - Image is then passed through optic nerve. Combined together for depth perception and to flip image. Then passed to the back of brain (visual cortex) via optic tract.
-
-
Edge Detection
-
-
This approximates the gradient of each pixel of the image, so the rate of change of colour/shades. Where there is a big difference, there could be an edge
Intensity changes - Depth, colour, texture, shadows, illumination, reflection
Smoothing - When noisy image is convolved with a smoothing function, the noise will be removed and any edges will be clearly shown when differentiated. Does widen the edge. Can skip a step by differentiating function then convolving with image
-
Canny - Use 1st order Gaussian and differentiate x and y with image. Then find magnitude, thin and threshold.
Thinning - Making thick lines in an image only 1 pixel wide. Basically finds the true edge and removes unnecessary data. Does this by looking and neighbour pixels of a given pixel and removing them if they are smaller, in x and y.
Hysteresis Thresholding - Uses a high and low threshold. High is the pixels that are almost definitely edges. Then uses the low to reconstruct the rest of edges, using directional information. Usually is half of high
ROC Analysis
-
-
-
-
-
-
When plotted, perfect result will be 1 TPR, 0 FPR.
Colour
Objects will absorb selective wavelengths and reflect others. So a red object absorbs all wavelengths in the visible light range, except red, which it reflects.
Black - Absorbs everything, emits nothing
White - Absorbs nothing, emits everything
Trichromatic Theory - We have 3 colours of cones in our eyes; red, green and blue. They are sensitive to those wavelengths of light. Explains differentiating colour shades, detect mixes of colours and colourblindness. Flaws with colour blending
Herring's Theory - Based on trichromatic. Two types of neurons; red-green and blue-yellow. If for eg light is detected from both green and red, neuron would not fire. So yellow would inhibit red-green and signal yellow from blue-yellow.
Represented as matrix. Colour images have 3 matrices, a red intensity matrix, green and blue. Then convoluted together
Hough Transform
Using a binary image, will look at every possible line that can be drawn over the image, and calculate the number of pixels that are there.
At a fixed point, lines are generated at the perpendicular of 0 < θ < 180 and length w.
-
-
Face Recognition
-
Create a co-variance matrix from a dataset of faces. Then calculate the eigenvectors of that matrix, which are the eigenfaces, which allows a face to be reconstructed.
-
Not all eigenfaces are needed to reconstruct a face. The faces with the highest weights will give a good outline of the face
Noise Filtering
-
Mean filter - uses an nxn matrix all of element m to convolve the image. If m is a fraction, it intensifies where there are bigger differential changes. Smoothing filter
Gaussian Filter - an nxn filter, where elements near the middle have a bigger weight. Values are calculated with a standard deviation