Target detection method based on multi-scale fusion lightweight deep learning convolutional network

Abstract

The invention discloses a target detection method based on a multi-scale fusion lightweight deep learning convolutional network. A backbone network is designed to be composed of two modules: a backbone feature extraction module and a multi-scale fusion positioning feature module; the trunk feature extraction module follows the structural feature that the DenseNet network is connected in series along the channel dimension, so that each layer is directly connected with all subsequent layers of the DenseNet network, features can be reused, redundant features do not need to be learned, the parameter quantity is reduced, and the network is kept simple and efficient; on the basis, a double-channel convolution channel mode is added, so that receptive fields of different scales are obtained; the multi-scale feature module continues to use an SSD multi-scale anchor frame detection mechanism, a 3-way residual module is added on the basis of the SSD multi-scale anchor frame detection mechanism, multi-scale features are fused, the expression ability of the features is enhanced, and therefore a multi-scale aircraft target is detected.

Description

technical field [0001] The invention belongs to the technical field of machine learning and deep learning, and in particular relates to a target detection method of a lightweight deep learning convolutional network based on multi-scale fusion. Background technique [0002] Usually, the number of parameters of a deep neural network can reach tens of millions or even billions. With the increase of the depth and breadth of the network, there are a large number of redundant parameters in the network model, which leads to a large amount of calculation and serious memory consumption of the deep learning model. main reason. In fact, in the process of network propagation, it is only necessary to retain neurons with high importance or organize representation neurons in a more optimized and simple way to provide sufficient feature information to the output layer of the neural network for result prediction. In response to this phenomenon, many methods for model compression and model o...

AI Technical Summary

Problems solved by technology

[0002] Usually, the number of parameters of a deep neural network can reach tens of millions or even billions. With the increase of the depth and breadth of the network, there are a large number of redundant parameters in the network model, which leads to a large amount of calculation and serious memory consumption of the deep learning model. main reason

Images