本课题围绕卫星侦察图像所具有的高清和稀疏特性，探索一种新型的低功耗模数混合级联型AI图像处理微系统架构及其电路实现。不同于传统的模拟CMOS传感器芯片与数字图像处理芯片所构成的处理系统，本课题提出了多级级联的模拟数字电路混合的图像处理结构。利用模拟电路的低功耗，高速特性和基于提前中断技术的数字图像筛选电路，将高清图像在CMOS图像传感器芯片内部进行预处理，仅将具有有效信息的少部分图像输出给后端AI处理器，降低运行复杂识别网络的后端AI处理器的输入图像大小，从而降低其运算量，以实现对高清侦察图像的实时本地化处理。相较于传统的图像回传地面做分析的处理方式，基于本地的AI图像处理系统可在灾害搜救时规避大量数据回传地面所带来的系统延时，在保证侦察范围和识别精度的同时，更加满足了救援的实时性。

This topic focuses on the high-resolution and sparse characteristics of high-altitude satellite reconnaissance images, and explores a new type of low-power analog digital hybrid cascaded AI image processing microsystem architecture and its circuit implementation. Different from the traditional analog CMOS sensor chip and digital image processing chip processing system, this topic proposes a multi-level cascaded analog and digital hybrid image processing architecture. Utilizing the low power consumption, high speed characteristics of the analog circuit and the early interrupt technology based digital image screening circuit, high-definition images are pre-processed inside the CMOS image sensor chip, and only a small part of the image with valid information is output to the back-end AI processor , Thereby reducing the input image size of the back-end AI processor, which makes real-time localizated processing of high-resolution reconnaissance images possible. Compared to the traditional image backhaul ground analysis method, the local AI-based image processing system can avoid the system delay caused by transfering a large amount of data back to the ground during disaster search and rescue, while ensuring the detection range and recognition accuracy. Thus, this system is more satisfied for the real-time requirement for rescue.