Herself’s Artificial Intelligence

Computers can usually out-compute the human brain, but there are some tasks, such as visual object recognition, that the brain performs easily yet are very challenging for computers. The brain has a much more sophisticated and swift visual processing system than even the most advanced artificial vision system, giving us an uncanny ability to extract salient information after just a glimpse that is presumably too fleeting for conscious thought. To explore this phenomenon, neuroscientists have long used rapid categorization tasks, in which subjects indicate whether an object from a specific class (such as an animal) is present or not in the image.

Now, in a new MIT study, a computer model designed to mimic the way the brain itself processes visual information performs as well as humans do on rapid categorization tasks. The model even tends to make similar errors as humans, possibly because it so closely follows the organization of the brain’s visual system.

More Information:
First Impressions: Computer Model Behaves Like Humans on visual Categorization Task

There are two types of computer eyes in use now. One known as ‘spot or jumping’ eyes send out a narrow laser beam and measure the amount of light that comes back. A second type of computer vision uses ‘imaging eyes’ these form pictures like the digital cameras do.

Image filtering from 2-d into 3-d involves filtering many items such as: shading; color; intensity; texture; etc. The information is then used to create a scene.

First the image is decomposed into pixels or small blocks of color and light intensity. Then to erase spikes of light or clean up random noise in the picture, a filter operation, convolution, is slid over the image that averages pixels that are adjacent to each other and replaces the center pixel with the average. Usually a two dimensional Gaussian is used as the weighting function. Another method looks for pixels that vary greatly from the pixels surrounding them. It gives this pixel the average of the surrounding pixels. This method tends not to blur the image as much as the first method.

Next some type of edge detection is performed. One method enhances areas that have large changes in color from pixel to pixel. This is done by a process that, in effect, takes the second derivative of the image. Where the second derivative is zero an edge occurs. This is done by creating a sliding window with a positive and negative section. As it slides over the image it compares the pixel underneath to itself. The sum is zero over areas that don’t change. A method used in maps is ‘contouring’ this measures the difference in intensity of the pixels rather than color. This pulls out and clarifies areas of different depths.

The convolution and edge detection are generally combined into one operation. This results in the Laplacian of the Gaussian function being performed over the image. There are newer, better techniques available now than this one.

Another method to differentiate objects from an image attempts to find areas of gradual change in color, light intensity, etc. An area is a region of homogeneous pixels that differ not more than a small amount, . Adjacent regions are not homogeneous. Split and Merge [Horowitz & Pavlidis] is one such method. The whole image is split into equal parts, these are tested for homogeneity, if the regions are not homogeneous then the splitting continues until all the regions are homogeneous. Regions are then merged with other regions that are homogeneous with themselves. This method and the one above run into many problems with differentiating shadows from edges.

Now scene analysis is done to extrapolate a scene from the information gathered. For this part more information is needed. Other scenes, stereo vision, or positions of the moving camera. In one method a line drawing is extrapolated and the junctions of the lines are matched to table entries to determine if the object extends outward or inward. If the scene contains well known objects the objects may be stored as line drawings in a table to be matched.