The term "Gestalt" denotes the perceived structure of visual input. (The analogous notion in language processing would be: the surface structure of utterances and texts.) We focus on static images. Their perception is determined by: (1) similarity with previously encountered images; (2) interpreting an image as a representation of a 3D scene, thereby accounting for perspective, lighting effects, etc.; (3) a preference for simplicity and regularity.

We now narrow our focus once more. We concentrate on abstract patterns where recognition and 3D-phenomena play no role. This is the domain of "Gestalt perception" in the more narrow sense of that term. (Note: It has been surmised that this exercise is intrinsically useless, because all perception is completely determined by the cognitive faculty of recognition. This "radical empiricism" has been tested and refuted. Cf. several papers in Ellis (1938).)
    

   
Max Wertheimer's Gestalt Principles

Wertheimer's thesis: Gestalt perception is determined by several interacting principles, such as:

Proximity:                       O  OO  OOO  OOO  OOOOOO   OO

Similarity in Color:            OOOOOOOOOOOOOOOOOOOO

Similarity in Size:             OOOooOoOOoooooOOOOOOooO

Similarity in Shape:           OOOXOOXXOOOOOOXOOXXXOX

Repetition:                      AXQAXQAXQAXQAXQAXQAXQAXQ

Symmetrie:                     BBBBBAAAXXXQQQXXXAAABBBBB

Experience:                    Recognizing previously perceived patterns

Proximity

Continuation

Similarity

Closure

References

Max Wertheimer: "Untersuchungen zur Lehre von der Gestalt." Psychologische Forschung 4 (1923), pp. 301-350. [English translation of important parts of this paper: "Laws of Organization in Perceptual Forms." In: W. Ellis: A source book of Gestalt psychology. London: Routledge & Kegan Paul, 1938, pp. 71-88.]

Irvin Rock and Stephen Palmer: "The Legacy of Gestalt Psychology." Scientific American, December 1990, pp. 48–61.

Emanuel Leeuwenberg: Structural Information Theory (SIT)

Domain: Sequences of symbols. (E.g.: "XPAPX", "LHOOQ", "OFHSKHKSHFKSHFKSH" – where the letters stand for atomic elements that have no perceivable relation to each other.)

The process of perception translates an input sequence into an expression of a "coding language". The coding language has the same primitive symbols as the input language, but it has other operations besides mere concatenation. These operations are: Repetition, Symmetry, and Alternation. E.g.:

Rep3(AB)           =  ABABAB
Sym(AB, C)        =  ABCBA
Altright(ABC, K) =  AKBKCK

Any input-pattern has many alternative codings. E.g.:

      ABABAB

= Rep3(AB)
= AB Rep2(AB)
= A Rep2(BA) B
= A Sym(BA, B)
= Altright(AAA, B)
= Altright(Rep3(A), B)

Perception is disambiguation. Disambiguation criterion: choose the shortest code. (Other psychologists have proposed similar ideas. Cf.: Fred Attneave, Max Clowes.) Note the similarity with the theory of Minimum Description Length (MDL), as developed by Kolmogorov, Solomonoff, Rissanen, Li & Vitanyi.

Reference

Emanuel L. J. Leeuwenberg: "A Perceptual Coding Language for Visual and Auditory Patterns." American Journal of Psychology 84, 3 (1971).
     

The discrete line: "turtle graphics"

Leeuwenberg's coding theory will not be the ultimate account of any domain of human perception. There are no perceptual modalities where one may realistically assume an input consisting of discrete, unrelated symbols. The visual domain presents an additional problem: it is intrinsically two-dimensional rather than linear. Leeuwenberg tries to sidestep the latter problem by focussing on black & white drawings consisting of straight line segments, and representing such drawings as discrete symbol sequences, by means of "turtle graphics".

In "Turtle Graphics", a line consisting of straight segments is described from the perspective of a turtle who draws it: the length of segment 1, the angle to turn after segment 1, the length of segment 2, the angle to turn after segment 2, and so on. Combined with the commands "pen up" and "pen down", this allows the description of all possible drawings consisting of straght line segments.

(A turtle graphics description specifies a sequence of points. Note that such a sequence of points can also be used to define a quadratic or cubic spline curve.)

References

Harold Abelson and Andrea diSessa: Turtle Geometry. The Computer as a Medium for Exploring Mathematics. Cambridge, MA: MIT Press, 1980.

Seymour Papert: Mindstorms. Children, Computers, and Powerful Ideas. New York: Basic Books, 1980.

Structural Information Theory for discrete line drawings

Using a turtle-graphics notation, the input-code of a square with edge a is:   pi/2 a  pi/2 a  pi/2 a  pi/2 a
The shortest SIT-coding of this pattern is:                                             rep4(pi/2 a)

A square with edge c with on its corners squares with edge a:                    rep4(pi/2 c rep4(pi/2 a))

This pattern is also perceived when the square with edge c is not drawn explicitly. To account for such examples, the operations "lift pen" and "lower pen" are added to the coding language.

The input code of a zigzag line:                 b a c a b a c a b a c a b a c a b a c a b a c a   

Its shortest coding in the SIT-language:      altright(rep6(b c), a)

A zigzag line is perceived as "symmetric" if c equals –b. The SIT-code does not account for this observation, because it ignores the relation between b and c: the angles are treated as atomic symbols.

Jos de Bruin's online SIT-Interpreter applet allows you to play with SIT-codes for discrete lines. [Note that this applet uses the SIT-notation defined in Chapter 2 of Dastani (1998), which is different than the one used in the examples above. Use the button "Auto" to see random examples of codes in that notation.]

Mapping arbitrary input-codes onto their shortest equivalent SIT-notation is an NP-complete problem. (The computational complexity is caused by the alternation-operation.) There are two implemented algrithms that try to meet this challenge.
Peter van der Helm's program PISA implements a deterministic algorithm which often (but not always) yields the optimal code. It was written in C and runs under Unix. Robert Voorn (VU) implemented a genetic algorithm which often (but not always) yields the optimal code. It was written in C++ and runs under Unix. Both programs can be obtained through Mehdi Dastani (mehdi at cs dot uu dot nl).

Dastani (1998) discusses some fundamental limitations of Structural Information Theory, and tries to overcome some of them. The most important limitations are:

References Gestalt Perception.

R. Collard, P. Vos and E. Leeuwenberg: "What Melody Tells about Metre in Music." Zeitschrift für Psychologie, 189 (1981).

Mohammad Mehdi Dastani: Languages of Perception. Ph. D. Thesis University of Amsterdam, 1998. ILLC Dissertation Series 1998-05.

Emanuel L. J. Leeuwenberg: "A Perceptual Coding Language for Visual and Auditory Patterns." American Journal of Psychology 84, 3 (1971).

---

References Complexity.

G. Birkhoff: "A Mathematical Approach to Aesthetics." Scientia 50 (1931), pp. 133-146.

R. Gunzenhäuser: "Birkhoff's Anwendungen." In: Maß und Information als Ästhetische Kategorien. Baden-Baden: Agis Verlag, 1975.

Frans Boselie and Emanuel Leeuwenberg: "A General Notion of Beauty Used to Quantify the Aesthetic Attractiveness of Geometric Forms." In: W. Crozier and A. Chapman (eds.): Cognitive Processes in the Perception of Art. Elsevier Science Publishers, 1984.

---

Acknowledgments

Several decades ago, Léon van Noorden drew our attention to Leeuwenberg's work.
Yuri Engelhardt provided some of the illustrations on this page.