Zero-Transition Serial Encoding for Image Sensors

Abstract

Off-chip serial buses are the most common interfaces between sensors and processing elements in embedded systems. Due to their length, these connections dissipate a large amount of energy, contributing significantly to the total consumption of the system. The error-tolerant feature of many sensor applications can be leveraged to reduce this energy contribution by means of an approximate serial data encoding. In this paper we propose one such encoding called Serial T0, particularly effective for image sensors data. By exploiting the correlation among pixels colors, Serial T0 significantly reduces the energy dissipation on a serial connection, with minimal impact on image quality, and extremely low codec overheads. Moreover, it allows runtime quality reconfiguration, and can be integrated with standard serial protocols (SPI, I2C, CSI-2, etc.) without requiring modifications to legacy peripherals. We show that Serial T0 is superior to both accurate and approximate state-of-the-art encodings, not only in terms of pure energy saving, but also when combined with realistic error tolerant applications (OCR and face recognition). In the best case, the energy on the bus can be reduced of more than 75% with less than 4% maximum error on the decoded data.

On the impact of smart sensor approximations on the accuracy of machine learning tasks

“…The energy benefits arising from signal approximation during data acquisition (described in Section 2.3 ) have an impact that depends on the nature of the signals and on the application which uses them. As observed in various previous works [7] , [8] , [9] , [10] , [11] , a common characteristic of sensor data is the burstiness of the signals, i.e. the fact that variations are concentrated in short time windows.…”

“…Moreover, alterations on a given block also affect downstream components. As a simple example, reducing the A/D bit resolution of an accelerometer from the typical 11- or 12-bit to 8-bit also speeds up and reduces the energy for data transmission [10] .…”

“…2 for three sensed signals: from and ECG sensor (a), from an accelerometer (b), and from an image sensor (c). An observation done in many previous works [8] , [10] , [11] is that the “relevant” data tend to be localized where variations occur. In the case of the ECG, spikes in the signal correspond to heart pulses, while almost-constant sections correspond to the interval between two heart beats.…”

“…Therefore, all subsystems are optimized for energy, including sensing, Analog-to-Digital (A/D) and Digital-to-Analog (D/A) conversion, processing, actuation and data transmission [1] , [5] . Such energy-optimized components often obtain high efficiency by means of various forms of data approximations [6] , [7] , [8] , [9] , [10] , [11] , [12] , [13] , [14] . They are, in other words, designed according to the so-called Approximate Computing paradigm, which has recently gained a lot of traction in both academia and industry [15] , [16] , [17] .…”

“…In this work we take on a different and novel perspective, namely, we assess the impact of the energy-driven approximations in the data acquisition path on the quality of the ML algorithms. Specifically, we focus on the approximated transmission of data from a smart sensor to a processor, through energy-efficient bus encodings [7] , [8] , [9] , [10] , [11] . Besides being relevant per se, due to the large amount of energy consumed by off-chip sensor-processor connections [10] , these approximations are also similar to (and therefore representative of) other data alterations that can appear in data acquisition chains, such as adaptive sampling frequency [6] , [30] , [31] and adaptive A/D conversion [32] , [33] .…”