• Tidak ada hasil yang ditemukan

Feature detectors

Dalam dokumen Professor Trevor Harley (Halaman 87-91)

If the presentation of a visual stimulus leads to detailed processing of its basic features, we may be able to identify cortical cells involved in such processing. Hubel and Wiesel (1962) studied cells in parts of the occipital cortex (at the back of the brain) involved with the early stages of visual processing. Some cells responded in two different ways to a spot of light depending on which part of the cell was affected:

1 An “on” response with an increased rate of firing when the light was on.

2 An “off” response with the light causing a decreased rate of firing.

Hubel and Wiesel (e.g., 1979) discovered two types of neuron in primary visual cortex: simple cells and complex cells. Simple cells have “on” and

“off” regions with each region being rectangular. These cells respond most to dark bars in a light field, light bars in a dark field, or straight edges between areas of light and dark. Any given cell responds strongly only to stimuli of a particular orientation. Thus, the responses of these cells could be relevant to feature detection.

Complex cells resemble simple cells in responding maximally to straight-line stimuli in a particular orientation. However, complex cells have large receptive fields and respond more to moving contours. Each complex cell is driven by several simple cells having the same orientation preference and closely overlapping receptive fields (Alonso & Martinez, 1998). There are also end-stopped cells. Their responsiveness depends on stimulus length and on orientation.

Weblink:

Hubel and Wiesel (1962)

IN THE REAL WORLD: FINGERPRINTING

An important form of pattern recognition in the real world involves experts matching a criminal’s fingerprints (the latent print) against stored fingerprint records. An automatic fingerprint identification system (AFIS) scans through huge databases (e.g., the FBI has the fingerprints of over 60 million individuals). This produces a small number of possible matches to the fingerprint obtained from the scene of the crime ranked in terms of similarity to the criminal’s fingerprint. Experts then decide which fingerprint (if any) in the database matches the criminal’s.

AFIS focuses on features at two levels. There are three general fingerprint patterns: loop, arch and whorl (circle). Fingerprints also contain more specific features. We have patterns of ridges and valleys known as friction ridges on our hands. Of particular importance are minutiae points – locations where a friction ridge ends abruptly or a ridge divides into two or more ridges. Experts are provided with information about feature or minutiae similarity from AFIS but also make use of microfeatures (e.g., the width of particular ridges).

You can see some of the complexities in fingerprint identification by deciding whether the two fingerprints shown in Figure 3.2 come from the same person. Four fingerprinting experts decided they came from the same person, namely, the bomber involved in the terrorist attack in Madrid on 11 March 2004. In fact, the fingerprints are from two different individuals. The one on the left is from the Madrid bomber (Ouhane Daoud), but the one on the right comes from Brandon Mayfield, an American lawyer who was falsely arrested.

Figure 3.2

The FBI’s mistaken identification of the Madrid bomber. The fingerprint from the crime scene is on the left. The fingerprint of the innocent suspect (positively identified by various fingerprint experts) is on the right.

From Dror et al. (2006). Reprinted with permission from Elsevier.

Findings

It is commonly believed that fingerprint identification is typically very accurate, with the case of the Madrid bomber a rare exception. In fact, that is not entirely the case. Cole (2005) reviewed 22 real-life cases involving fingerprint misidentifi cation by experts. In over half the cases, the original expert misidentification was confirmed by one or more additional experts. Dror et al. (2012) had experts list all the minutiae on ten fingerprints and then repeat the exercise a few months later. There was total agreement between their two assessments of the same fingerprint only 16% of the time.

Many mistakes are made because of the intrinsic complexity and incom pleteness of latent fingerprints. However, top-down processes also contribute towards identification errors. Many such errors involve forensic confirmation bias, which Kassin et al. (2013, p. 45) defined as “the class of effects through which an individual’s pre-existing beliefs, expectations, motives, and situational context influence the collection, perception, and interpretation of evidence”.

Dror et al. (2006) reported evidence of forensic confirmation bias. Experts judged whether two fingerprints matched having been told incorrectly that the prints were the ones mistakenly matched by the FBI as the Madrid bomber. Unknown to these experts, they had judged these fingerprints to be a clear and definite match several years earlier. The misleading information provided led 60% of the experts to judge the prints to be definite non-matches! Thus, top-down processes triggered by contextual information can distort fingerprint identification.

Further evidence of forensic confirmation bias was reported by Langenburg et al. (2009). They studied the effects of context (e.g., alleged conclusions of an internationally respected expert) on fingerprint identification. Experts and non-experts were both influenced by contextual information but non-experts were influenced more.

Why are experts better than non-experts at deciding accurately that two fingerprints match? According to signal-detection theory (e.g., Phillips et al., 2001), there are two possibilities.

First, experts may have an excellent ability to discriminate between prints that match and those that do not match. Second, they have a lenient response bias, meaning that they have a strong tendency to respond “match” to every pair of prints regardless of whether there is actually a match. The acid test of which explanation is more applicable is the false-alarm rate – the tendency to respond “match”

incorrectly to similar but non-matching pairs of prints. Good discrimination is associated with a low false-alarm rate whereas a lenient response bias is associated with a high false-alarm rate.

Thompson et al. (2014) carried out a study on experts and novices using genuine crime scene prints. Both groups responded “match” accurately on approximately 70% of trials on which there was a genuine match. However, there was a substantial difference in the false-alarm rate. Novices incorrectly responded “match” when the two prints were similar but did not match on 57% of trials compared to only 1.65% for experts. These findings indicate that experts have much better discrimination than novices. They also have a much more conservative response bias than novices, meaning they are more reluctant to respond “match”.

All these types of cells are involved in feature detection. However, we must not exaggerate their usefulness. These cells provide ambiguous information because they respond in the same way to different stimuli. For example, a cell may respond equally to a horizontal line moving rapidly and a nearly horizontal line moving slowly. We must combine information from many neurons to remove ambiguities.

Hubel and Wiesel’s theoretical account needs to be expanded to take account of the finding that neurons differ in their responsiveness to different spatial frequencies (see later section entitled “spatial frequency”). As we will see, several phenomena in visual perception depend on this differential responsiveness.

PERCEPTUAL ORGANISATION

It would probably be fairly easy to work out which parts of the visual information available to us belong together and thus form objects if those objects were spread out in space. Instead, the visual environment is often complex and confusing, with many objects overlapping others and so hiding parts of them from view. As a result, it can be hard to achieve perceptual segregation of visual objects.

The first systematic attempt to study these issues was made by the Gestaltists. They were German psychologists (including Koffka, Köhler and Wertheimer) who emigrated to the United States between the two World Wars. The Gestaltists’ fundamental principle was the law of Prägnanz, according to which we typically perceive the simplest possible organisation of the visual field.

KEY TERM

Law of Prägnanz

The notion that the simplest possible organisation of the visual environment is perceived; proposed by the Gestaltists.

Most of the Gestaltists’ other laws can be subsumed under the law of Prägnanz. Figure 3.3(a) illustrates the law of proximity, according to which visual elements close in space tend to be grouped together. Figure 3.3(b) shows the law of similarity, according to which similar elements tend to be grouped together.

We see two crossing lines in Figure 3.3(c) because, according to the law of good continuation, we group together those elements requiring the fewest changes or interruptions in straight or smoothly curving lines. Finally, Figure 3.3(d) illustrates the law of closure: the missing parts of a figure are filled in to complete the figure (here a circle).

Weblink:

Gestalt laws of perceptual organisation

Weblink:

Article by Max Wertheimer

Figure 3.3

Examples of the Gestalt laws of perceptual organisation: (a) the law of proximity; (b) the law of similarity; (c) the law of good continuation; and (d) the law of closure.

As Wagemans et al. (2012a) pointed out, it is easy to dismiss these grouping principles as “mere textbook curiosities”. In fact, however, the various grouping principles “pervade virtually all perceptual experiences because they determine the objects and parts that people perceive in their environment”

(Wagemans et al., 2012a, p. 1180).

The Gestaltists emphasised the importance of figure–ground segregation in perception. One part of the visual field is identified as the figure, and the remainder is treated as less important and so forms the ground.

KEY TERMS

Figure–ground segregation

The perceptual organisation of the visual field into a figure (object of central interest) and a ground (less important background).

According to the Gestaltists, the figure is perceived as having a distinct form or shape whereas the ground lacks form. In addition, the figure is perceived as being in front of the ground and the contour separating the figure from the ground is seen as belonging to the figure. You can check the validity of these claims by looking at the faces–goblet illusion (see Figure 3.4). When the goblet is perceived as the figure, it seems to be in front of a dark background. Faces are in front of a light background when forming the figure.

Various factors determine which region is identified as the figure and which as the ground. Regions that are convex (curving outwards), small,

surrounded and symmetrical are more likely to be perceived as the figure than regions lacking these characteristics (Wagemans et al., 2012a). For example, Fowlkes et al. (2007) studied numerous natural images for which observers made figure–ground decisions. Figure regions tended to be smaller and more convex than ground regions. Overall, the findings indicate that cues emphasised by the Gestaltists are, indeed, important in figure–ground assignment.

Figure 3.4

An ambiguous drawing that can be seen as either two faces or as a goblet.

Weblink:

Figure–ground segregation

Dalam dokumen Professor Trevor Harley (Halaman 87-91)