Object Detection Based On Depth Learning For Multi-Feature Fusion

Object detection is the process of identifying a target and locating it in a given image or video frame.It is a prerequisite for many computer vision tasks and is also widely used in practical tasks.However,the traditional method of object detection has spent a lot of time in the design of features,and the features of manual design are not universal,resulting in low detection accuracy.Now often through deep learning methods,the use of convolutional neural networks instead of hand-designed features.However,existing object detection methods often use pre-training models to improve accuracy,and in some scenarios,pre-training models can not be used.If we can have a network structure,we can use random initialization method to train the network and obtain satisfactory detection accuracy,which is of great significance in application.This paper presents a multi-feature fusion object detection method that has a simple training method and can improve the detection accuracy.In order to solve the problem that the pre-training model can not be used in some scenes,this paper designs a neural network structure that can not use the pre-training model.In the training process,the parameters of the neural network model are initialized by random initialization and then loaded with a large number of pictures and data sets constructed by the corresponding tags,the pictures in the data set are preprocessed,the data are expanded,and the tagging information and the cost function are iteratively updated Neural network model parameters until convergence;in the prediction process,based on the trained neural network model,the image or video frame to be predicted is taken as an input,the target is identified and its positioning is given.In order to verify the accuracy of the proposed algorithm,the object detection experiments were carried out in the PASCAL VOC dataset,and the average accuracy of the proposed algorithm was averaged in each image,reaching an average detection rate of 78.8%.In the same case without using the pre-training model,the mean average detection rate is slightly higher than the current optimal model,and the detection speed is about three times.The network structure designed in this paper can detect 20 categories of targets simultaneously with PASCAL VOC dataset training.Experimental results show that the convolution neural network model based on feature fusion is faster than most object detection models and the object detection accuracy is higher.Finally,thanks to the advantage of network structure,The method of training achieves the object detection model that does not lose through the transfer learning training.