Real-time video recognition accelerator architecture based on key object splicing

Abstract

The invention discloses a real-time video recognition accelerator architecture based on key object splicing. The real-time video recognition accelerator architecture comprises an object tracking module, an object aggregation module, an object splitting module, a preset neural network accelerator, an object queue updating module and a main storage module. The object tracking module is used for acquiring original position information of a key object rectangular frame in P-frame image data or original position information of a key object rectangular frame in B-frame image data; the object aggregation module is used for merging key object rectangular frames in the P-frame image data and / or key object rectangular frames in the B-frame image data to obtain a synthetic frame; the preset neural network accelerator is used for processing the composite frame to obtain a composite frame identification result; and the object splitting module is used for splitting the composite frame and returning a splitting result to the original image data. According to the architecture, the calculation workload in a target video recognition task is greatly saved, and the recognition task processing speed and the recognition accuracy are improved.

Description

technical field [0001] The invention relates to the technical field of neural networks, in particular to a real-time video recognition accelerator architecture based on key object stitching. Background technique [0002] Deep convolutional neural networks have been widely used in image recognition, such as in image classification, detection and segmentation. With its development, people have gradually extended the application range of deep convolutional neural networks to the video field. [0003] Usually, video recognition tasks based on deep neural networks can treat each video frame as an independent picture and input it into the deep neural network for recognition, that is, video recognition is regarded as an image recognition task to recognize each frame separately. However, directly applying a network model suitable for image recognition tasks to all video frames requires a huge computational and energy overhead; on the other hand, neural networks applied to image rec...

AI Technical Summary

Problems solved by technology

[0005] The technical problem to be solved by the present invention is that it is often necessary to process each frame of video image in the existing neural network processing video recognition task, which has the disadvantages of long time-consuming, high energy consumption, difficulty in guaranteeing recognition accuracy and difficulty in improving recognition speed.

Images