Model Lightweight Method for Object Detection

Abstract

The rapid development of object detection technology benefits from the development of convolutional neural network. However, the convolution neural network needs a deep enough convolution layer to obtain more abundant image feature information and the complexity of the network itself, which makes the object detection network have some limitations, such as large amount of model parameters, unable to achieve real-time detection speed, high requirements for computing resources and so on. Based on the efficientdet model and the LD-BiFPN network, this paper explores the difference between the adaptive fusion method and the fast fusion method, and designs a feature layer pruning method of the weight matrix according to the weight matrix representing the importance of the feature layer in the fast fusion method, to prune the LD-BiFPN network to reduce the network parameters. Then convolution filter pruning method is introduced to prune the classification and regression network of the model, so as to reduce the parameters of the network and improve the detection speed. Experiments show that the designed lightweight method can reduce the parameters of the model and improve the detection speed on the premise of ensuring the detection accuracy.t