Aircraft Image Management Based on AI Typical Mainland Zynq

In the aircraft management system, the universality of aircraft target detection neural network model compression has been verified for many engineering application conditions such as full-range intelligent maneuvering target recognition in specific areas, long-distance maneuvering target detection, and deep space recognition of non-cooperative targets. A data set suitable for aerospace has been selected to complete variable-scale recognition without basically reducing the accuracy. The model is compressed to one-tenth, and the inference time is controlled at the millisecond level. Using neural network compression and quantization after pruning, a 90% compression rate SSD fixed-point model is achieved to control the resize process and effectively prevent precision dissipation, target loss of lock, and energy consumption expansion. Multi-acceleration IP core parallel computing is deployed on the mainland Zynq series single chip to achieve a peak performance of 1.2TOPS and an mAP value of 0.95, meeting the strict requirements of aerospace vehicles for single-time target recognition accuracy.