Research On Target Detection Based On Deep Full Convolution Network

With the rapid development of Internet technology,target detection tasks are closely related to people's daily life.This task has been widely used in many fields such as road monitoring,medical diagnosis and video information retrieval.Therefore,image target detection technology has become a popular research direction in the field of computer vision.In recent years,the emergence of deep learning has made the research in this field a big step.The target detection framework of deep learning has achieved better detection results.The convolutional neural network relies on its powerful feature extraction ability,which can guarantee Adapting to more complex scenarios with high accuracy,current mainstream algorithms can be classified into two categories: target detection algorithms based on candidate regions,which have problems such as slow detection speed and cumbersome process,and another regression-based algorithm.The target detection algorithm is at the expense of precision,which reduces the practicability of the algorithm.Therefore,this paper proposes an improvement scheme based on the problems existing in the algorithm and the advantages of the regional and non-regional target detection algorithms.The main work of the thesis is as follows:(1)Aiming at the problem of detecting the flaws in the complex background and the small-scale target process,the candidate region-based full convolution network target detection algorithm RFCN is improved,and a multi-scale feature fusion model is proposed.The model uses the reverse connection method to fuse the multi-layer convolution features in the network,and uses the target prior to reduce the sample search space.Finally,it is better to describe the target by combining the low-level high-resolution features and the high-level rich semantic information.Convolution feature.The experimental results show that the detection model of fusion convolution features effectively improves the detection ability of complex background targets and small-scale targets,and at the same time considers the speed of target detection.(2)Aiming at the problem of regression location of candidate region frames,this paper proposes a twodimensional loss function to reduce the loss between the prediction candidate frame and the real bounding box to ensure accurate positioning of the target.The experimental results show that the loss function effectively achieves the accurate positioning of the target,which will be more helpful for the detection of small targets.(3)For the problem that the proportional imbalance between the positive and negative samples in the training process leads to the slow convergence of the model training and the weak generalization ability of the model,the online model integration optimization training is proposed by using the online difficult sample mining algorithm.The algorithm can balance the proportion of positive and negative samples used in training,speed up the convergence of the model and make network training more adequate.After experimental verification,the improved network model detection accuracy has been significantly improved,and to some extent,the problem of small target easy to miss detection has been solved.