Algorithm Research On Target Detection Based On Fully Convolutional Networks

With the explosive growth of image data,image processing has become more and more important.Significant target detection is one of the important research directions of image processing.Target detection is widely concerned by researchers because of its large number of application requirements in the real world,such as automatic driving,video surveillance,robot vision,etc.In recent years,the high accuracy demonstrated by deep learning to solve complex problems has made it widely used and studied in target detection.Compared with the traditional target detection algorithm,the target detection algorithm based on the deep convolution neural network relies on the neural network when extracting the target features.The position and type of the target can be inferred by the obtained features,so as to improve the accuracy of detection.The core task of this paper is to study the existing target detection algorithm,analyze the advantages and disadvantages of the algorithm,design an improved algorithm to improve accuracy and practicability,and verify the accuracy and practicability of the model through experiments.(1)The research background and significance of target detection are introduced,and the development status of convolution neural network and target detection technology are analyzed.At the same time,the convolution neural network,dilated and deconvolution are introduced in detail.The segmentation technology used in target detection is also summarized from semantic image segmentation technology.The advantages and disadvantages of segmentation technology are briefly introduced.These research contents lay a theoretical foundation for multi-scale integration and time cost optimization.(2)Aiming at the problem of ambiguous boundary of target detection based on fully convolutional networks,this paper summarizes the two jump connection structures before,and adopts a new multi-scale feature fusion fully convolutional networks for target detection.Firstly,through jump connection,deep feature and shallow feature are combined at channel level.The boundary information is added,and then the fused feature map is convoluted to further extract features.(3)The superposition of multi-layer convolution layers will result in large computational complexity,increased time cost and limited pixel receptive field.Aiming at the problem of stacking multi-layer convolution layers,this paper designs a fully convolutional networks model based on high-efficiency pyramid convolution network and jump connection.The shallow feature map is convoluted with voids with different expansion rates,and then integrated according to certain rules to further extract depth features.The efficient pyramid convolution network continuously reduces the amount of computation and increases the receptive field.(4)Experiments on the ade20 k dataset by scene analysis show that the training loss of the fully convolutional networks model with new jump connection structure converges faster and the verification loss is smaller.The detection accuracy results based on the efficient pyramid convolution network and the fully convolutional networks model with jump connections are 0.73% higher than the recent advanced algorithms,which increases the number of convolution layers while ensuring the stability of computation time cost.