Research On General Object Detection Method Based On Deep Learning

Object detection technology,as one of the core research topics in the field of computer vision,is an important computer vision task.In recent years,with the improvement of hardware computing capabilities,the birth of large data sets and the development of deep learning technologies,the performance of target detection has been greatly improved.Among them,due to its structural advantages,deep convolutional neural networks have incomparable advantages in traditional object detection in image feature extraction,powerful feature extraction capabilities and learnability,and can effectively extract image feature information.In the existing target detection methods,detection accuracy and detection speed cannot always be taken into account,and how to achieve accurate and accurate target detection is still a major problem for researchers.Aiming at the problems such as the inaccuracy and speed in the existing target detection framework and the poor detection of small targets,this paper proposes a new type of convolutional neural network based on the target detection algorithm based on the candidate region idea.The main research contents are as follows:(1)Because the traditional CNN network model has many parameters and high feature map dimensions,it results in large memory consumption,large amount of data,high calculation cost,long detection time,and it is also difficult to deploy on mobile devices with limited memory resources.on.Therefore,in order to reduce the parameters in the feature extraction stage,this paper adds a deep separable convolution layer on the backbone network for feature extraction to compress the feature map,thereby reducing the amount of data and computing cost.Depth separable convolution can solve the standard convolution integration into deep convolution and point-by-point convolution to complete the extraction and combination of image features in two steps,which can reduce the amount of parameters and calculation cost of the model.The experimental results show that after increasing the depth separable convolutional layer,the parameters of the model are greatly compressed,the memory consumption is reduced,and the detection speed is improved.(2)At the same time,in order to avoid the loss of accuracy caused by the amount of parameters and further increase the detection accuracy of the network,this paper adopts a key point detection method based on a full convolutional network.The main two-step operation of target detection method is to classify and locate the target.Traditional target detection methods use regression-based methods to locate the target,map the feature map into a vector through the fully connected layer,and the method used in this paper divides the candidate area of the object into some key points,and applies a full Convolutional neural network(FCN)is used to predict the position of grid points.Due to the position sensitivity of the full convolutional network structure,this method guarantees clear spatial information,and the positions of key points can also be obtained at the pixel level After obtaining a certain number of points at the feature position,the corresponding candidate frame is also determined.Therefore,the key point-oriented method can obtain more accurate target candidate frames than the traditional regression-based method.(3)In addition,for the problems of easy to miss detection and low detection accuracy when detecting small targets,this paper uses the PS-Rol Align method to replace the traditional pooling method to improve the entire detection system's perception of small targets..At the same time,a multi-scale feature fusion model is also proposed in this paper.The model uses reverse connection to fuse multi-layer convolutional features in the network,and uses a multi-scale RPN network to finally obtain convolutional features that are more conducive to the description of the target by combining low-level resolution features and high-level rich semantic information.Through experimental verification,the model after convolutional features is integrated,which effectively improves the detection ability of small targets,while taking into account the accuracy of object detection to a certain extent.