Embedded Image Recognition System for Lightweight Convolutional Neural Networks

Abstract

In this paper, we design and implement an embedded image recognition system based on STM32 for the problem of limited storage space of embedded systems to run convolutional neural networks efficiently, and for the loading of lightweight convolutional neural network and the hook-up requirement of the quadrotor. The system hardware adopts the idea of modular design to improve the compatibility of the system, and the system software adopts the training of handwritten image recognition based on convolutional neural network, lightweight processing of the convolutional neural network, and transplanting the trained network to the embedded system. Finally, the system can finish the recognition of handwritten images stably and efficiently. This system can provide a low-cost and highly integrated solution for such image processing systems. The lightweight target detection model CED-Det is designed by combining CED-Net and dense feature pyramid network, which firstly performs feature extraction by CED-Net, then performs feature fusion by stacking two layers of dense pyramid network, and finally, the fused feature maps are used for classification prediction and position prediction by two 3×3 convolutions, respectively. CED-Det is used in VOC and Experimental results on COCO datasets show that CED-Det is more suitable for embedded platforms in terms of accuracy, inference speed, and a total number of parameters compared with other target detection models.