Object Recognition and Visual Processing

Low-Level Processing:
- Extraction of basic visual features like orientation, color, contrast, disparity, and movement.
- Early visual processing areas (e.g., LGN, V1) are involved.
- Receptive fields help identify bars and edges (orientation selectivity in V1).
Intermediate-Level Processing:
- Combination of low-level features to build more complex representations.
- Identification of contours, surface properties, and shape discrimination.
- Segregation of objects based on depth information.
High-Level Processing:
- Identification of objects (e.g., a horse) in a viewpoint-invariant manner.
- Requires making sense of all extracted information to recognize objects regardless of perspective.

Goal: To identify a particular object and build a representation that corresponds only to that object.
Steps:
- Edge detection: Early receptor fields (retina, LGN) with center-surround organization.
  - Photoreceptor signals combine to create excitatory (ON) or inhibitory (OFF) responses.
  - Example: A light bar stimulating the ON region results in a burst of action potentials.
- Orientation selectivity: Combining LGN neuron responses in V1 to achieve orientation-specific responses.
  - Example: Three LGN neurons with receptive fields oriented vertically combine to create a vertically oriented response in a cortical cell.
- Curvature detection:
  - V4 neurons code for complex image properties in terms of surface shape.
  - Neurons are tuned to specific curvatures, which is useful for detecting complex shapes.
- Combining simple cells from V1 with different orientation preferences can build curvature-selective responses in V4.

LGN:
- Circular center-surround receptive fields.
- Respond to changes in light intensity.
  $Equation:$\ \text{Excitation} \rightarrow \text{Action Potentials}
  $Equation:$\ \text{Inhibition} \rightarrow \text{Reduction in Action Potentials}
V1:
- Orientation-selective cells (simple cells).
- Combine LGN neuron responses.
- Respond to bars and edges of specific orientations.
  $Equation:$\ \text{Vertically Oriented Light Bar} \rightarrow \text{Strong Response}
  $Equation:$\ \text{Change Orientation} \rightarrow \text{Reduced Response}
V4:
- Cells code for more complex image properties like surface shape and curvature.
- Combining simple cells from V1 can build curvature-selective responses in V4.
  $Equation:$\ \text{Curve Providing Light Stimulus} \rightarrow \text{Excitatory Response}
  $Equation:$\ \text{Move Curve to the Left} \rightarrow \text{Inhibition}

What pathway: Ventral stream.
IT cortex critically involved in neural responses that correspond to familiar objects.
Hierarchy of processing: V1 -> V2 -> V4 -> IT.
Low-level structural representations in V1.
Mid-level structural representations (contours) in V2/V4.
High-level structural representations (object recognition) in IT.
Synthesis of information about form, color, and depth.
Neurons in IT respond poorly to simple stimuli (spots, lines).
Large receptive fields, integrating information from larger retinal regions.
Responses tend not to change when an object moves or changes size within the receptive field (viewpoint invariance).

Visual object agnosia: Loss of ability to recognize familiar objects through vision.
Inferotemporal cortex active when humans look at objects vs. scrambled images.
Pattern of activation in IT determines what objects are perceived.
Different types of agnosia depending on the location of damage.
- Can recognize and order shapes but can't identify same shapes.
- Inability to copy or understand letters and their relation.
Recognition loss is modality-specific.
- Can identify a key by touch but not by sight.

Structural Mechanisms:
- Based on identifying features associated with an object (e.g., a chair has four legs and a back).
Holistic Processing:
- Recognizing objects by processing all features simultaneously in the correct configuration (e.g., face recognition).
Two types of agnosia:
- One for structural mechanisms.
- One for holistic processing (face processing).
Faces as special stimuli:
- Small changes in the configuration of facial features allow us to identify individual faces.
- Even when faces are constructed from other objects (e.g., vegetables), the configuration can trigger face recognition.

Faces are typically processed in an upright configuration.
Less sensitive to overall configuration when faces are inverted.
Inversion effect indicates that faces are special and different from other objects.

Recording from single cells in the inferior temporal cortex.
Neurons respond to images of faces.
Greatest responses to faces with the correct configuration of features (eyes, nose, mouth).
Less response to scrambled or obscured faces.
Responses are typically greatest for images that look most like faces.

fMRI shows greater response to faces compared to scrambled images in a region of the temporal cortex.
Fusiform face area (FFA) is particularly responsive to faces.
Some debate about whether FFA is specific to faces or is an expert object recognition area.
- Some evidence suggests that FFA might be activated when looking at images of things we might be an expert looking at.
- More recent work suggests the FFA is strongly associated with faces.
Adjacent brain regions respond to objects, faces, body parts, and scenes.

Results in the inability to recognize faces.
Normal visual recognition of other objects.
The individual can still recognize people through other modalities (e.g., voice).

Good ability to recognize and remember scenes.
Good at identifying different scene categories.
Able to distinguish between a large number of examples or contexts of a given scene category.
Scene recognition might be a little like faces.
Natural scenes can be recognized at a high fidelity.

Parts of the brain used for processing scenes and objects are not the same.
Scene perception operates independently of object perception.
Patient DF: Profound visual agnosia for objects but able to recognize scenes.
Parahippocampal place area (PPA): A brain region important for recognizing different places.