A deep neural network object detection method based on feature multiplexing

Abstract

The invention provides a deep neural network target detection method based on feature multiplexing, which has a clear network structure, simple training algorithm, can significantly reduce the amount of network parameters, and can also maintain high detection accuracy. It includes steps: S1: generate target candidate frames of different shapes and different proportions centered on each pixel of the feature map, and obtain the feature map to be classified; S2: construct the target detection framework; S3: train the target detection framework, and obtain the trained Target detection framework; S4: input the feature map to be classified obtained by the target candidate frame in step S1 into the trained target detection framework for classification; S5: judge the feature map to be classified by the classification result obtained in step S4 is the background or is the target to be tested, and then the detection of the object is realized by calculating the object category and position offset of the feature map to be classified.

Description

technical field [0001] The invention relates to the technical field of image recognition, in particular to a deep neural network object detection method based on feature multiplexing. Background technique [0002] In the field of computer vision, typical target detection frameworks mainly use deep neural networks as the basic network, and the recognition accuracy is relatively ideal; however, most of these networks in the prior art are training networks based on large data sets, with large parameters At the same time, the existing target detection framework mainly focuses on the improvement of detection accuracy and detection speed, while ignoring the optimization of network parameters; a large number of redundant parameters bring a large consumption of computing resources and time costs , thus reducing the network distributed training efficiency and increasing the data transmission burden, making it difficult for these networks to achieve real-time object detection on embed...

AI Technical Summary

Problems solved by technology

[0002] In the field of computer vision, typical target detection frameworks mainly use deep neural networks as the basic network, and the recognition accuracy is relatively ideal; however, most of these networks in the prior art are training networks based on large data sets, with large parameters At the same time, the existing target detection framework mainly focuses on the improvement of detection accuracy and detection speed, while ignoring the optimization of network parameters; a large number of redundant parameters bring a large consumption of computing resources and time costs , thus reducing the distributed training efficiency of the network and increasing the burden of data transmission, making it difficult for these networks to achieve real-time object detection on embedded devices with limited computing resources

Images