Research On Object Detection Algorithm Based On Deep Learning

Object detection is an important branch in the field of computer vision.Its main purpose is to analyze and process a given video or picture,indicate the category to which each object belongs,and draw a bounding box near the object to mark the position of the object.Traditional object detection algorithms are based on manually extracted features.The algorithms have low accuracy and weak generalization performance.In recent years,deep learning technology has developed rapidly.As deep learning-based object detection algorithms use convolutional neural networks to extract features,they have made breakthrough progress in detection accuracy and are widely used in video surveillance,smart logistics,medicine image analysis,driverless and other fields.However,in actual application scenarios,the speed of object detection is slow,the detection effect is often affected by the deformation,occlusion and environmental changes.Designing a good real-time and robust object detection algorithm has very important research significance.This paper proposes corresponding solutions to the problems which exist in deep learning-based object detection algorithm,and it has a reference significance for the study of deep learning-based object detection algorithms.The main research work and research results of this article are as follows:1)A multi-scale feature detection method based on clustering ideas is proposed.Most object detection algorithms usually detect the last layer of feature maps extracted by the network,and the network parameter settings are relatively arbitrary.In order to solve this problem,this paper adopts a multi-scale feature fusion detection method by detecting objects on feature maps of different scales to increase the number of candidate frames of the network.When setting the network parameters,the size of the "anchor" is designed by clustering the data set in advance,which speeds up the training process of the network.Finally,it is verified on the general object detection data set,and it proves the necessity and feasibility of multi-scale feature detection method based on clustering idea in object detection algorithm.2)A real-time object detection algorithm based on residual network is designed.The detection speed of the target detection algorithm based on deep learning is often related to the computing power of the hardware.In this regard,this paper designs a realtime object detection algorithm based on residual networks.By combining the structure of residual networks and deep separable convolutions,the network model is reduced while ensuring good detection performance of the network model.The amount of parameters improves the detection speed of the algorithm.The verification on the universal detection data set proves that the algorithm proposed in this paper meets the real-time requirements in terms of detection speed.3)For the problem of low detection accuracy of real-time object detection algorithm,two optimization methods are proposed.Generally,when the algorithm's detection speed is optimized,the detection accuracy will often decrease.In this paper,in order to solve the problems of uneven distribution of data samples in the real-time target detection algorithm and the bad effect of occlusion target detection,a new crossentropy loss function and an improved non-maximum suppression algorithm are designed respectively.The new loss function not only reduces the impact of the uneven number of samples in different categories on the accuracy of object detection,but also dynamically adjusts the proportion of simple samples and difficult samples that affect the loss function,it enhancing the algorithm's detection accuracy on difficult samples.The soft-NMS algorithm,when dealing with overlapping objects,reduces the probability that the candidate frames around the objects are suppressed,thereby improving the detection recall index.Finally,it is verified by experiments that the accuracy of the optimized real-time object detection algorithm has been significantly improved.