A Lightweight Deep Learning Target Detection Algorithm Based On Multiscale Feature Fusion

Convolutional neural networks are widely used in the field of computer vision.In order to improve the accuracy of convolution neural network,the network structure increases the number of parameters and gradually increases the network complexity,making it difficult to run smoothly in real time on small and medium platforms(such as mobile and embedded devices).Lightweight convolution neural networks are better suited to specific application scenarios where memory resources are low,processor performance is low,and power consumption is limited.This thesis studies how to improve the target detection accuracy of the Lightweight Convolution Neural Network under the premise of simple and efficient,as follows:1.Before the output of the final feature layer of the Lightweight Convolution Neural Network,do feature fusion between the shallow and deep levels,and propose a feature fusion structure suitable for the Lightweight Network to enhance the semantic information with fewer parameters.2.On the pooling layer,the pooling layer is used as an important structure to fuse the different scale features of a single layer,and the different scale features of the same layer are fused into the same layer to enhance the features.3.Simply change the structure on the basic component unit of the classical lightweight convolution neural network,introduce coordinate attention mechanism,and use the new unit as the main structure of the basic network for light feature extraction,make more use of the feature information at all levels of the lightweight convolution neural network,reduce the loss of features,and improve the ability of feature representation.In this thesis,a new type of lightweight convolution neural network composed of feature extraction basic network,feature enhancement module and feature fusion module is presented.The experimental results on VOC dataset are greatly improved without expanding the computational cost and number of parameters,which is of practical significance.