Broadband convolutional processing is critical to high-precision image recognition and is of use in remote sensing and environmental monitoring. Implementing in-sensor broadband convolutional processing using conventional complementary metal–oxide–semiconductor technology is, however, challenging because broadband sensing and convolutional processing require the use of the same physical processes. Here we show that a palladium diselenide/molybdenum ditelluride van der Waals heterostructure can provide simultaneous broadband image sensing and convolutional processing. The band alignment between type-II and type-III heterojunctions of the photovoltaic heterostructure is gate tunable, and the devices exhibit linear light-intensity dependence for both positive and negative photoconductivity, as well as linear gate dependence for the broadband photoresponse. Our in-sensor broadband convolutional processing improves recognition accuracy for multi-band images compared with conventional single-band-based convolutional neural networks.

宽带卷积处理对于高精度图像识别至关重要，可用于遥感和环境监测。然而，使用传统的互补金属氧化物半导体技术实现传感器内宽带卷积处理具有挑战性，因为宽带传感和卷积处理需要使用相同的物理过程。在这里，我们表明二硒化钯/二碲化钼范德华异质结构可以同时提供宽带图像传感和卷积处理。光伏异质结构的 II 型和 III 型异质结之间的能带排列是栅极可调的，并且器件对正和负光电导率都表现出线性光强度依赖性，以及宽带光响应的线性栅极依赖性。与传统的基于单波段的卷积神经网络相比，我们的传感器内宽带卷积处理提高了多波段图像的识别精度。