Deformable convolutional network operation method, device and system based on FPGA

Abstract

The invention provides a deformable convolutional network operation method, a device and a system based on an FPGA, and the method comprises the steps: obtaining image input data, and obtaining serialdata after the caching processing; performing matching segmentation on the serial data according to the convolution size to obtain a plurality of to-be-convolved fragments; and according to a presetdeformable convolution kernel added with an offset vector and the weight corresponding to each to-be-convolved segment, performing convolution operation on each to-be-convolved segment in sequence toobtain a convolution output result. Through the deformable convolution algorithm, the operation of any convolution model is realized, the image detection precision of the convolution module can be effectively improved, and meanwhile, a better hardware acceleration effect is obtained by virtue of an FPGA platform.

Description

technical field [0001] The present invention relates to the technical field of convolutional neural networks, in particular to an FPGA-based deformable convolutional network operation method, device and system. Background technique [0002] Image classification has always been the core problem to be solved in the field of computer vision. The convolutional neural network is a multi-layer perceptron with powerful feature learning and abstract expression capabilities. The multi-layer network structure can directly use images as network input. Through training Image features are obtained, which simplifies the feature extraction process. However, the usual convolutional network structure has a high recognition rate for text and simple graphics, but its ability to classify large data and complex images is insufficient, and overfitting also leads to weak generalization ability. [0003] Each logic module of FPGA is independent of each other, and it has strong parallel processing ...

Images