An improved Tiny-YOLOv3 and its implementation with FPGA

Tiny-YOLOv3 is a simplified YOLO algorithm and has the characteristics of simple network model and small computational cost, which is very suitable for real-time target detection applications. Aiming at low accuracy of Tiny-YOLOv3 used in detecting small target objects, Tiny-YOLOv3 algorithm is improved by changing two-scale detection to three-scale detection and calculating the parallelism ratio of the loss function with CIoU (Complete-IoU) in this paper. Meanwhile, the BN (Batch Normalization) layer is merged into the convolutional layer, which speeds up the forward inference of Tiny-YOLOv3. The improved Tiny-YOLOv3 is realized in an FPGA (Field programmable gate array) to detect targets. The experimental results show that detection accuracy of the improved Tiny-YOLOv3 is increased by 48.6% and the detection speed of the improved Tiny-YOLOv3 is decreased by only 5% compared with Tiny- YOLOv3. It is suitable for realization on FPGA.