In-sensor Synaptic Computing in a Curved Image Sensor via Three-Dimensional Heterogeneous Integration

Image sensor technology provides a bridge between the physical world of analogue image/light signals and virtual digital data by converting light signals to electrical signals. There has been a dramatic improvement in silicon-based image sensor technology, in which the number of pixels and nodes has been exponentially increased in virtue of the rapid development of semiconductor fabrication technology. When a large amount of raw data generated in the photodetectors are transported to the computing system through limited sensory nodes, the conventional image sensors, however, suffer from a data bottleneck effect at the sensor/processor interface as well as within the processor with the von Neumann computing architecture, resulting in data transportation and computation delay. This delay can be a critical issue to image sensor applications that require fast image processing and strict delay requirements, such as autonomous driving and real-time video analysis. Furthermore, a planar geometry of the typical image sensors still needs bulky and complex optics lenses to minimize spherical aberration. The bulky lens system is the main obstacle in further miniaturization of the planar image sensors. In this proposal, I will develop a curved computational image sensor with the capability of in-sensor computing via heterogeneous integration and thin-film III-V compound semiconductor fabrication techniques. This entirely novel image sensor system will effectively overcome the abovementioned limitations in conventional image sensors by adopting the hemispherical geometry of the imager and in-sensor computing architecture.