Research On Small Object Detection Method Based On Convolutional Neural Network

In this era of rapid development of information technology,target detection has become an important direction in the field of visual research,and at the same time,it embodies its important value in many fields.In complex environments,the target is interfered by factors such as occlusion and illumination,which leads to increasing requirements on the model.Therefore,it is necessary to overcome more difficulties and disturbances to improve its robustness.How to design a model that can accurately identify the target under the influence of various factors has become the top priority in this field.The early algorithm combining sliding window and image zooming not only has high detection cost but also low efficiency.Most of the efficient algorithms are proposed based on convolutional neural network(CNN).Although the overhead of the algorithm is controlled to a certain extent and the detection accuracy of general objects is improved,the effect is still not ideal in the detection of the small-size object s.It is found that the feature map in the lower layer of the network has a high resolution but insufficient abstract semantics,and the feature map in the higher layer has a small size due to multiple down-sampling operations but rich semantic information,resulting in the loss of information related to edge details and target positioning.When the actual size of the target is small,there is a greater demand for the expression ability of the feature.The feature graph at the lower level is sampled more precisely to make up for the feature of the object with deeper semantics.For this reason,this paper proposes a method of full connection between densely connection blocks to fuse the features of different levels,and to precisely sample the features of shallow layers,so as to generate more feature semantic information.The research contents of this paper are as follows:(1)In view of the existing algorithm to extract features of the small-size object inadequate,easily lost that the small object characteristics in the part of the scenario of image,thus the problem of the low of detection accuracy.In this paper,a method of densely and fully connection between pieces of fusion characteristics,at the level of implementation of each layer characteristic information reuse,to a certain extent,make up for the model to extract feature information is missing,has been more robustness to the improvement of the characteristics of the testing accuracy.(2)Reference residual structure of ideas,design to increase the similar structure of convolution and deconvolution on super-resolution technique in network.Through the short connection by leaps and bounds to achieve the purpose of correction proposal position,strengthen the performance model of object location,speed up the convergence speed of the Network,and RoI Align to avoid losses,area don't match how to reduce the demand for computing power,improve the model on the small object detection accuracy.(3)For medium and small target image boundary ambiguity,background and noise ratio is big,it is difficult to small-size object detection problem.Draw lessons from the idea of SENet,increase the MSC module to optimize the extraction features,characteristics which network of the final output,at the same time,increase attention machine in the channel and spatial on the dimension to further enhance the detection precision.In addition,the experiments in this thesis all used the recognized data set,designed the ablation experiment comparison,and verified the effectiveness of the improved model for the detection of the small-size object.Our method is more robust in feature extraction and has higher detection quality.When the mAP reaches 81.6%,AP(such as bottle,etc.)increases to 72.4%,which has obvious advantages over other representative algorithms.