Design and implementation of a hardware accelerator IP core for improved lightweight deep learning model

Abstract

Real-time multi-point, full-scene monitoring with low cost, low power consumption, low communication overhead, and front-end deployment is a current research focus in fire detection technology. This paper investigates and implements fire detection technology on the low-computation ZYNQ platform based on deep learning, aiming to provide a cost-effective, highly efficient, and reliable fire detection solution. Firstly, we propose a lightweight network model, YOLO-Fire, which incorporates modifications like replacing standard convolutions with depthwise separable convolutions, adding the ECA attention mechanism, and introducing multi-scale feature fusion to suit the memory and computational limitations of the ZYNQ device. Additionally, we designed a hardware accelerator IP core for the ZYNQ7020 platform using a specific loop tiling strategy, constraint statements, and a dual-dimensional parallel optimization of convolution input and output channels. Combined with fixed-point quantization and resource optimization, this implementation achieves efficient acceleration of convolution, pooling, and upsampling layers. Experimental results show that YOLO-Fire improves accuracy, recall, and F1-score on the BoWFire public flame dataset and a self-constructed flame dataset. Additionally, the average inference time on the ZYNQ platform is approximately 74.43 times faster than on mainstream ARM AI platforms, verifying the effectiveness of the proposed acceleration approach.