Abstract: SAR image target recognition mainly aims at strategic military targets such as bridges and airports, as well as tactical targets such as aircraft, tanks, and automobiles. Accurate identification, classification and positioning is an important part of SAR image interpretation. Firstly, the main processing layer of the convolutional neural network based on C6678 is constructed. Secondly, the processing and storage characteristics of C6678 are combined to optimize the design of the convolutional layer and network scheduling. Finally, a design and implementation method of the YOLOv3-TINY target recognition network on C6678 is completed. This method can reconstruct and modify common convolutional neural network models, and solves the problem of running deep learning networks on multi-core processing platforms such as C6678. The experimental results show that the method is consistent with GPU in detection performance. Considering the real-time image frame rate of airborne SAR, although the real-time performance of this method on C6678 is far from that of GPU, it can meet the real-time processing requirements of airborne SAR.

Key words: synthetic aperture radar(SAR), target recognition, you only look once(YOLO), DSP；deep learning