PaperlayerView exact revision on arXiv

Source-linked AI summary

STEMPix: A Phase-Transition-Material-Based Pixel Sensor for Resolving Edge-Movement Direction

Md Rahatul Islam Udoy, Sumeet Kumar Gupta, Deep Jariwala, Ahmedullah Aziz

arXiv:2609.04435v1cs.ETcs.CV

TL;DR

Full-frame intensity readout and external processing can be costly when local boundary movement is the information of interest. STEMPix combines temporal pixel changes with neighboring-pixel spatial edges to produce local direction codes. Its estimated implementation reports compact pixel pitch and low active switching energy.

  • Problem

    For edge-movement sensing, full multi-bit frame readout and external processing impose bandwidth costs, while compact local operations must avoid unnecessary full-image processing.

    Source: §I · p. 1, §II · p. 4

  • Method

    STEMPix combines temporal changes in binarized pixel values with neighboring-pixel spatial edges to generate 3-bit local edge direction codes.

    Source: §I · p. 3, B. LEDC Generation and Interpretation · p. 6

  • Results

    0.465 fJ is the estimated average active switching energy per LEDC operation, and the estimated implementation has a 1.73 μm × 2.36 μm pixel pitch and 95.47% geometric fill factor.

    Source: B. Switching Energy Analysis · p. 12, §VIII · p. 13

  • Takeaways & Limitations

    The results indicate that local edge-movement coding can provide a compact front-end representation for specialized sensing where edge movement is task-relevant.

    Source: §I · p. 1, §VIII · p. 13

    STEMPix is intended for specialized sensing where local edge movement matters more than full-scene intensity reconstruction, not as a general-purpose camera replacement.

    Source: §I · p. 1

Abstract

from arXiv · show

This paper proposes a spatio-temporal edge-movement direction pixel (STEMPix) for generating compact direction-aware edge movement information inside a CMOS-compatible image sensor array. The proposed design targets specialized sensing applications where local boundary movement is more important than full-frame intensity reconstruction. Instead of transferring full multi-bit frames for external processing, STEMPix generates a 3-bit local edge direction code (LEDC) by combining pixel-level temporal change information with neighboring-pixel spatial edge information. We design the architecture using a two-tier organization, where the photodiode layer is separated from the computation layer to preserve light-collection area while accommodating the additional in-array processing circuitry. The proposed circuit is evaluated through HSPICE transient simulations. The estimated implementation achieves a horizontal pitch of 1.73 {\mu}m, a vertical pitch of 2.36 {\mu}m, and a geometric fill factor of 95.47%. The average active switching energy is 0.465 fJ per LEDC operation across representative edge-movement cases. The proposed STEMPix operation also supports global-shutter capture and dynamic thresholding. These results indicate that STEMPix can provide a compact and scalable front-end representation for edge-movement-aware sensing systems.

I. INTRODUCTION

Local edge movement can represent object-boundary changes when full image intensity is not the task-relevant information. STEMPix combines temporal pixel changes with spatial edge information to generate compact edge-direction codes inside the array.

I. INTRODUCTION

  • Local edge movements can indicate object changes in applications where boundary motion matters more than full-frame intensity.Examples include expanding fire fronts, growing cracks, moving blast boundaries, and changing contours.Source: §I · p. 1
  • Reading full multi-bit frames for external processing creates bandwidth, power, and latency costs, while direct in-pixel processing can increase pixel pitch and power.The design challenge is to select compact operations that avoid reconstructing or processing the full image.Source: §I · p. 1
  • STEMPix combines per-pixel temporal events with neighboring-pixel spatial edges to generate a 3-bit local edge direction code inside the array.The design is evaluated with HSPICE transient simulations and layout-based footprint estimation; the sensing and computation layers can be separated in a stacked organization.Source: §I · pp. 1, 3
  • STEMPix targets specialized sensing rather than general-purpose image reconstruction, and its timing scheme is designed to support global-shutter-style operation.The paper notes that pixels should make binary decisions for the same time instant before sequential readout begins.Source: §I · p. 3

II. DESIGN MOTIVATION AND TECHNICAL BACKGROUND

The section motivates detecting object changes from local edge movements and contrasts that approach with full-frame processing. It introduces PTM-based HyperFET thresholding as the means of binarizing pixel signals.

II. DESIGN MOTIVATION AND TECHNICAL BACKGROUND

  • Object translation, shrinking, and expansion can be inferred from the directions of local boundary movements between consecutive frames.Diagonal movement can be decomposed into horizontal and vertical edge movements.Source: §II · p. 3
  • External edge-movement analysis requires reading full multi-bit frames, whereas the proposed array locally generates a compact representation of motion-related changes.The conventional pipeline sends analog pixel signals through column circuits and processes the resulting frames externally.Source: §II · p. 4
  • In a conventional 3-transistor active pixel sensor, the photodiode discharges according to incident light during integration, and the source follower and selector support readout.An NMOS reset charges the photodiode node to approximately VDD - VTH; a PMOS reset can reset it to VDD.Source: §II · p. 4
  • A PTM-based HyperFET provides sharp switching for threshold-based pixel binarization as the PTM undergoes an insulator-to-metal transition at critical voltage and current.The PTM is integrated at a MOSFET source terminal; its critical parameters may shift at elevated temperatures and can be engineered for a targeted high-temperature range.Source: §II · pp. 4–5

III. STEMPIX ARCHITECTURE AND OPERATION

This section introduces the proposed STEMPix architecture and its circuit-level implementation.

III. STEMPIX ARCHITECTURE AND OPERATION

  • The section presents the STEMPix architecture and its circuit-level implementation.Source: §III · p. 5

A. Block Level Overview

STEMPix combines temporal changes in individual binary pixel values with spatial edge information between neighboring pixels. Its temporal event map marks changing locations regardless of object-background contrast polarity.

A. Block Level Overview

  • An event means a binary pixel-value change between consecutive frames; a local edge means neighboring pixels currently have different binary values.When objects move, binary values change near their boundaries; the same event-map pattern holds when object and background contrast are reversed.Source: A. Block Level Overview · p. 5
  • EventPix cells detect frame-to-frame binary pixel changes, while spatial BGD blocks detect edges between neighboring pixels.Together, these outputs provide temporal and current-frame spatial information for coding local edge direction.Source: A. Block Level Overview · p. 5
  • The temporal event map is identical for bright-on-dark and dark-on-bright scenes because events mark binary-value changes, not contrast polarity.The map captures where pixel values changed between consecutive frames.Source: A. Block Level Overview · p. 5

B. LEDC Generation and Interpretation

LEDC encodes local edge movement at a neighboring-pixel boundary using two temporal event bits and one current-frame spatial edge bit. The same three-bit structure applies to horizontal and vertical pairs, with code positions mapped to the pair’s orientation.

B. LEDC Generation and Interpretation

  • For a horizontal pair, LEDC is [TEL, SE, TER]: the outer bits report left- and right-pixel changes, and the center bit reports a current spatial edge.In Fig. 4, these positions come from OUTTE(1,1), OUTSE, and OUTTE(1,2), respectively.Source: B. LEDC Generation and Interpretation · p. 5, Figure 4 · p. 6
  • For a vertical pair, the same code positions represent top-pixel change, the edge, and bottom-pixel change, respectively.The left and right bit names refer to code positions, not physical locations, for vertical pairs.Source: B. LEDC Generation and Interpretation · p. 5
  • With a spatial edge present, 011 indicates leftward movement, 110 rightward movement, 010 a stationary edge, and 111 an edge-flipped condition.For vertical pairs, 011 and 110 indicate upward or downward movement according to the selected top-to-bottom orientation.Source: B. LEDC Generation and Interpretation · pp. 5–6
  • LEDCs are assigned to boundaries between neighboring pixels, so one pixel can participate in multiple overlapping horizontal and vertical codes.The code represents local spatio-temporal behavior around a pixel pair rather than the state of one isolated pixel.Source: B. LEDC Generation and Interpretation · pp. 5–6
  • Three bits are the minimum needed to encode whether each neighboring pixel changed and whether a current-frame edge exists between them.Temporal bits alone cannot confirm an edge, while one temporal bit plus the edge bit cannot identify which side changed.Source: B. LEDC Generation and Interpretation · p. 6

C. Circuit Configuration

STEMPix combines per-pixel temporal-event generation with neighboring-pixel spatial comparison to produce a 3-bit local edge direction code. Its circuit thresholds photodiode responses into binary values, compares current and previous values and demonstrates codes for representative edge movements.

C. Circuit Configuration

  • Each EventPix stores current and previous binary pixel values, while a neighboring-pixel BGD compares current values to detect a local binary edge.An XOR generates each pixel’s temporal event; another XOR implements the illustrated BGD comparison.Source: C. Circuit Configuration · pp. 6–7

D. Circuit Working Mechanism

  • At threshold time Tth, the HyperFET reference VREF converts photodiode voltage into a binary value: Crpix is low if Vph falls below VREF and high otherwise.A higher light intensity discharges Vph faster; the threshold decision is latched for the current cycle.Source: D. Circuit Working Mechanism · p. 7
  • The first-cycle temporal-event readout is invalid until Crpix is transferred to Prpix, which supplies the previous-cycle value for the next comparison.The temporal XOR output is high when current and previous binary values differ.Source: D. Circuit Working Mechanism · p. 7
  • The three simulated cases produce LEDCs 011 for an edge moved left, 111 for an edge flipped, and 010 for a stationary edge.The codes combine temporal events from two adjacent pixels with the spatial edge signal between them.Source: D. Circuit Working Mechanism · p. 7, Figure 5 · p. 8
  • The same LEDC mechanism applies to vertical as well as horizontal boundaries by changing the relative pixel positions and selected neighboring-pixel boundary.For a horizontal pair, the code uses the two pixels’ temporal-event outputs and the spatial edge between them.Source: D. Circuit Working Mechanism · pp. 7, 9

IV. ARRAY-LEVEL DESIGN AND OPERATION

The array shares control signals and collects temporal-event and spatial-edge outputs through dedicated lines. Reset, integration, thresholding, and storage occur across the array together, while readout proceeds sequentially to form horizontal and vertical LEDCs.

IV. ARRAY-LEVEL DESIGN AND OPERATION

  • EventPix cells occupy rows, with horizontal BGDs between same-row neighbors and vertical BGDs between neighboring rows.Control signals and supply rails are shared along each row, and the schedule alternates horizontal and vertical operations.Source: §IV · pp. 9–10
  • All pixels perform reset, light integration, thresholding, and pixel-value storage simultaneously before row-wise readout, supporting global-shutter-like operation.The schedule separates globally performed capture and storage phases from sequential readout.Source: §IV · pp. 9–10
  • Horizontal LEDCs use two temporal-event outputs and their same-row spatial edge; vertical LEDCs select adjacent rows and use odd-row and even-row event lines.Separate parity lines keep the two pixels’ temporal-event outputs distinct during vertical readout.Source: §IV · p. 10
  • The estimated two-layer layout places the photodiode and reset transistor in the first layer and the remaining computation circuits in the second.A 0.4 μm × 0.4 μm hybrid-bonding landing region connects the layers.Source: §IV · p. 10

A. Footprint

The footprint estimate uses a two-tier layout that keeps the photodiode in the sensing layer and places most computation circuitry in the second layer. The estimated pitches and geometric fill factor are derived from a 14 nm FinFET layout.

A. Footprint

  • The two-tier estimate places the photodiode, reset transistor, and hybrid-bonding landing region in the first layer, with remaining processing circuits in the second.The second layer accommodates thresholding, latching, pixel-value storage, temporal-event generation, and spatial-edge generation.Source: A. Footprint · pp. 10–11, Figure 7 · p. 11
  • The pitch estimate uses 14 nm FinFET technology parameters and derives the horizontal and vertical dimensions from the EventPix and spatial BGD layout.The hybrid-bonding landing region is added laterally to the horizontal pitch, not the vertical pitch.Source: A. Footprint · p. 10, Figure 7 · p. 11
  • The estimated horizontal pitch is 1730 nm, vertical pitch is 2360 nm, and geometric fill factor is 95.47%.The estimate subtracts reset-transistor and hybrid-bonding landing-pad areas from the available first-layer photodiode area; it assumes backside illumination and excludes optical effects such as microlens focusing.Source: A. Footprint · p. 10, Table II · p. 11

B. Switching Energy Analysis

The switching-energy analysis estimates active energy over representative LEDC-generation cases using HSPICE transient simulation. The reported average active switching energy is 0.465 fJ.

B. Switching Energy Analysis

  • For each case, active switching energy is summed over supply sources and selected intervals, excluding long inactive intervals.The calculation multiplies average supply power by each active interval’s duration, then averages across cases.Source: B. Switching Energy Analysis · p. 11
  • 0.465 fJ is the estimated average active switching energy per STEMPix operation across representative LEDC-generation cases.The cases include edge movement left, right, edge flipping, and a stationary edge.Source: B. Switching Energy Analysis · pp. 11–12

C. Frame Time

Frame time combines globally performed operations with sequential horizontal and vertical LEDC readout. The section also verifies adjustable binary decisions and global-shutter storage that remains fixed after latching.

C. Frame Time

  • (4Nrow + 11.7) ns gives frame time, which scales linearly with the number of array rows.The calculation adds 13.7 ns before first-row readout to 2 ns for each of 2Nrow − 1 readout steps.Source: C. Frame Time · p. 12

VI. DYNAMIC THRESHOLDING AND GLOBAL-SHUTTER OPERATION VERIFICATION

  • Changing thresholding time changes the stored binary pixel decision even under fixed illumination.The photodiode voltage falls over time: at Tth1 the pixel stores Crpix = 1, while at Tth2 it stores Crpix = 0.Source: §VI · p. 12
  • After thresholding is latched, Crpix remains fixed despite subsequent photodiode discharge below VREF.Global thresholding and latching let all pixels store decisions at the same selected time for later temporal-event generation, spatial-edge generation, and sequential LEDC readout.Source: §VI · p. 12

VII. BENCHMARKING

STEMPix produces a compact local edge-movement code by combining temporal pixel changes with neighboring-pixel spatial edges. Its estimated implementation has a 95.47% geometric fill factor and 0.465 fJ average active switching energy per LEDC operation.

VII. BENCHMARKING

  • Unlike edge-extraction sensors that process signals in readout circuitry and event sensors that output temporal activity, STEMPix locally combines temporal events and spatial edges into a direction-aware code.The comparison is intended to position STEMPix's functionality relative to the closest available sensor architectures.Source: §VII · pp. 12–13, Table III · p. 13
  • STEMPix's reported energy is active switching energy for local LEDC generation, so comparisons should account for differences in the compared sensors' metrics.The passage notes that prior edge-extraction works use larger technology nodes and column/readout-level processing, while STEMPix uses an estimated 14 nm FinFET implementation with local binary operations.Source: §VII · p. 13

VIII. CONCLUSION

  • 1.73 μm × 2.36 μm pixel pitch, 95.47% geometric fill factor, and 0.465 fJ average active switching energy per LEDC operation are reported for the estimated implementation.The conclusion attributes the compact structure to a two-tier organization that moves selected processing closer to the sensing plane while maintaining most sensing-layer area for photodiode integration.Source: §VIII · p. 13
Loading 2609.04435v1…