Embedded platform face detection method based on octave convolution and YOLOv3

Abstract

The invention provides an embedded platform face detection method based on octave convolution and YOLOv3. The embedded platform face detection method comprises the following steps: S1, collecting a video stream and transmitting a video frame; S2, supplementing edges of the frame picture; S3, zooming the picture; and S4, sending the picture into a detection end-to-end model based on YOLOv3 and octave convolution, merging results and performing filtering. According to the octave convolution and YOLOv3-based embedded platform face detection method provided by the invention, the forward compatibility of the latest network module under a caffe framework is ensured; the maintenance cost of different embedded platforms on the algorithm is reduced, the model deployment is convenient, a tedious cascade architecture is abandoned, an end-to-end architecture capable of directly obtaining a result in forward calculation is used, and the development difficulty is greatly reduced.

Description

technical field [0001] The invention belongs to the field of computer vision, and in particular relates to an embedded platform face detection method based on octave convolution and YOLOv3. Background technique [0002] At present, the mainstream method of face detection is the face target detection method based on MTCNN or FasterRcnn in deep learning. Both of them are cascaded structures in terms of neural network structure. MTCNN is composed of three parts: PNet, RNet, and ONet, and is given by PNet. A large number of target candidate frames, RNet and ONet are filtered separately to obtain the final result. The difference of FasterRcnn is that the Rcnn part maps a large number of candidate frames given by RPN to the original image, and then performs weighted screening. During the deployment process on the embedded platform, the cascaded structure may bring a large amount of computing consumption, and frequent switching between the CPU and the embedded core intensifies the ...

AI Technical Summary

Problems solved by technology

During the deployment process on the embedded platform, the cascaded structure may bring a large amount of computing consumption, and frequent switching between the CPU and the embedded core intensifies the waiting time of the algorithm, and due to the inherent problem of the method, a model with a large amount of parameters must be used to achieve a better performance. effect, which also increases the memory usage of the chip

Images