Sparse method for target detection network and accelerator design

• Main Authors: CHO SEONG-HO, LIU HAO, HOU ZONGHAO, WU ZHONGXING
• Format: Patent
• Language: Chinese; English
• Published: 14.07.2023
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; PHYSICS

The invention discloses a sparsification method and accelerator design for a target detection network, the network is sparsified for a YOLOv2-tiny target detection network under the condition of ensuring certain precision, the network has hardware friendliness, and the designed accelerator can efficiently support network model operation after sparsification, so that the network model operation is more accurate. The designed calculation unit has two calculation modes, and the network sparsity brought by the sparsification method can be fully utilized. The core of the rarefaction method is to take a convolution kernel as a unit, remove a value with a smaller absolute value, encode a reserved non-zero value position, and provide the encoded non-zero value position for a calculation unit to match feature map data with a weight during calculation, so that a large amount of redundant calculation is reduced, the storage space is saved, the operation speed is improved, and the calculation efficiency is improved. And