Research On Fast Target Detection Algorithm Based On Deep Learning

Object detection,a basic research hotspots in the field of computer vision,serves as a prerequisite for a large number of advanced visual tasks and also plays an important role in face recognition,intelligent monitoring,humancomputer interaction and other fields.Many traditional object detection methods are based on manual feature,and the detection performance of these algorithms is easily affected by the factors of environment lighting,occlusion,and the object deformation,etc.With the rise of deep learning,researchers have combined object detection tasks with deep neural networks in recent years,ma king deep learning surpass traditional visual methods with great advantages.Features learned on the big data in deep learning are far superior to the traditional manual features in terms of performance.While having better performance than the traditional ones,deep learning based object detection methods also have obvious disadvantages such as complexity of the network model,difficult to train,huge amount of calculation and poor real-time performance.These disadvantages brings many difficulties when applying the object detection algorithm to the actual scene.In order to solve this problem,this paper proposes an end-to-end fast object detection method by fusing low-level features.The specific research content is as follows:(1)Research on object detection algorithm based on deep learning.We study in-depth the basic structure and principles of classical convolutional neural network.Target detection based both on proposal regions and end-toend method are realized,and the detection results are compared and analyzed.(2)Aiming at the problem of low precision of SSD(Single Shot MultiBox Detection)algorithm and its insufficient ability to detect small targets,a lowlevel feature fusion detection model is proposed.Specifically,based on the structure of SSD network,the improved model generates feature maps by fusing the third and fourth layers of the base network and adjusts the resolution by the Atrous convolution layer.The resulting feature maps could provide the features needed for the small target detection.In addition,we also designed a set of effective data augmentation methods in the data preprocessing stage.Experiment results show that compared with the original SSD detection framework,the imp roved algorithm in th is paper has higher detection accuracy and more robust,and the detection effect on small targets is improved significantly.