Object Detection Algorithm Based On Improved Residual Network And Multi-scale Feature Fusion

Object detection is the premise of many computer vision tasks,which has been widely used in face recognition,automatic driving and other fields.Object detection can detect the specific location and type of objects from the image.With the development of artificial neural network,the traditional object detection algorithm has been gradually replaced by the object detection algorithm based on neural network,the detection accuracy and speed are constantly improved and the practicability is constantly enhanced.Therefore,it is of great significance to study the object detection algorithm based on neural network.The main work of this paper is to improve the defect of CenterNet object detection algorithm.The mian works are as follows:Firstly,a feature extraction backbone network based on the residual network improvement is proposed to solve the problems of network structure bloat and feature loss when feature extraction is carried out in CenterNet.After comparing the performance of the residual network with different layers,the 34-layer residual network with more balanced detection speed and precision is selected as the improved basic network.In order to solve the problem of feature loss in the process of feature extraction,the skip architecture is used to fuse the details and spatial information of the shallow layer with the deep semantic information.In order to solve the problem of weak generalization ability caused by shallow network,a strategy of network stack is proposed to deepen the network structure.Secondly,in view of the problem that CenterNet can’t effectively use multi-scale features in the network by using single scale feature detection.A multi-scale feature detection method is designed: according to the idea of feature pyramid network,multiple feature maps with different scales are selected in the improved residual network.We design a multi-scale feature maps fusion module to fuse the selected multi-scale feature maps into a feature map.The object boxs are predicted on the fused feature map.Thirdly,the defect of loss function in the training is improved.L1 loss is used in the loss function of the network.The derivative of the loss function is large when the error is small,which is not conducive to the convergence of the network,and when the error value is 0,it is not differentiable.It is proposed to use Huber Loss to replace and can successfully make up for the defects of L1 Loss.Finally,the comparison experiment proves that the improved feature extraction backbone network,the multi-scale feature detection method and the improved loss function have significantly improved the detection accuracy and speed of CenterNet on the PASCAL VOC dataset.