A Real-Time Memory-Less In-Sensor Time-Domain Convolution Processor with Programmable Kernel for Feature Extraction

With the growing demand for artificial intelligence (AI) and the Internet of Things (IoT), there is an increasing need for smart vision sensors and cameras with energy-efficient computing capabilities. While previous works have explored in-sensor computing for feature extraction, they often rely on memory-based weight storage or fixed kernels, limiting their flexibility and energy efficiency. This paper introduces a low-power, programmable, memory-less convolution engine designed for feature extraction in the analog domain. The proposed engine utilizes a linear large-signal voltage-to-current converter-based Time-Domain (TD) multiply-accumulate (MAC) cell, employing time pulses as weights. A 32-phase subsampling phase-locked loop (SS-PLL) is implemented to generate 5-bit Time Domain weights for a programmable 3x3 kernel employed for feature extraction. The in-sensor convolution engine achieves an energy efficiency of 0.4 pJ/pixel at a data rate of 300 MSps, making it suitable for resource-constrained, battery-operated devices.