Research On Multi-Object Detection Based On Deep Learning And Its Application

In this paper,multi-object detection based on deep learning has carried out a series of research in data security,multi-scale object detection,rotating object detection,and so on.The main work and contributions include the following:1)Aiming at the problems of data security and data island in object detection,a model optimization method based on vertical federated learning is proposed.Firstly,aiming at many invalid parameters in the model,a solution of parameter sharing,and gradient compression is proposed;Secondly,aiming at the problem of training speed reduction caused by sharing encryption parameters between communication parties,the parameters of communication parties are modified based on homomorphic encryption.This method not only protects the datasets of both sides of communication,solves the problems of data security,and data island,but also reduces the number of model parameters and speeds up the speed of model training and model reasoning.2)Aiming at the difficulties of small object detection and the imbalance between positive and negative samples in object detection,an object detection algorithm based on depthwise separable residual network is constructed.On the one hand,the feature extraction network model is modified,and the standard convolutional neural network structure is modified to a depthwise separable residual network.On the other hand,the loss function is modified,and the object box size and box position loss function is modified from mean square error to CIoU,and the confidence loss function is increased with positive and negative sample control parameters.The model obtained by training with this algorithm better solves the problems of small object detection difficulties and positive and negative sample imbalance.3)Aiming at the problems of different illumination intensities in object detection,such as inaccurate model recognition,difficult small object detection,and multi-scale detection,an object detection algorithm based on multi-scale fusion is constructed.The feature extraction network CourNet is proposed,and the Focus module is used to slice the image data on the basis of CourNet.We perform multi-scale feature fusion based on PANet.The method not only solves the problems of multi-scale detection and the difficulty of small object detection,but also improves the detection accuracy of objects under different illumination intensities.4)Aiming at the problems of poor real-time detection of rotating objects and inaccurate identification of dense objects in object detection,a fast and lightweight rotating object detection algorithm is proposed.The network is composed of a backbone,four scale fusion network,and rotating branches.Firstly,a lightweight network unit SLeanNet is designed,and this unit is used to build a low-cost and accurate backbone network.Then,a four-scale feature fusion module is designed to generate a four-scale feature pyramid,which contains more abundant ship shapes and texture features,and is conducive to the detection of small ships.Finally,we design a novel rotating branch module,which uses balance L1 loss function and R-NMS for post-processing to realize the accurate positioning and regression of rotating frames in one step.The algorithm improves the accuracy of fast detection of rotating object frames.