Research And Design Of Object Detection Based On Convolution Neural Network

With the advent of the era of big data and the development of smart phone technology,the image data on the Internet shows explosive growth.Image data is uploaded to the Internet every day.How to efficiently dig the information of image date has become a challenge.As example,some image data needs to locate and identify the objects for extracting the information.Therefore object detection become a research hotspot.In recent years,with the continuous development of deep learning technology,the object detection algorithm based on convolutional neural network has made higher detection accuracy than traditional methods.But new method also brings some new problems,such as small object detection,overlapping object detection and so on.In view of these new detection problems,this paper starts from designing of the loss function of the detection region,analyzing and studying the loss functions of others' object detection algorithm,and Combining these different loss functions in Cascade?RCNN network model to improve detection accuracy.According to the specific analysis and experiment of various detection region loss functions,this paper discusses the advantages and disadvantages of various detection region loss functions and the standard to measure which detection region loss function is better.According to this,several detection region loss functions based on different measurement standards are proposed in this paper.Cascade?RCNN object detection network which uses Res Ne XT-101 as the backbone network is used to verify and analyze the improvement of different loss function.Finally,two detection region loss functions COLloss and CLloss are proposed,and the feasibility of COLloss and CLoss is verified by experiments.At the same time,the standard of measuring the detection region is applied to the non maximum suppression algorithm,and a new non maximum suppression algorithm based on the distance of the center point is proposed.Then the algorithm which combined the traditional non maximum suppression based on IOU and the new non maximum suppression algorithm base on the distance of the center point is proposed,and the feasibility of the algorithm is verified by experiments.To dig the detection potential of Cascade?RCNN.This paper propose a new method which Combining a variety of different detection region loss function and putting it into Cascade?RCNN's unique multiple detection box regression module to improve the detection accuracy.Through experiments and analysis find that the combination of loss function which can improve the detection effect need the optimization direction of each loss function in the combination to be consistent and similar.Then comparing and analyzing the train time and the total loss function of different combination of region loss function.Further,comparing and analyzing the result of different combination of loss function on small object detection,overlapping object.Taking these as evaluation criterion to select the best combination of loss function which is COL+S+S+SThe experimental results show that the COL+S+S+S detection frame and CL+S+S+S detection frame proposed in this paper can effectively detect the target.The m AP of the original Cascade?RCNN on the Pascal?VOC data set is 0.826,and the m AP of the COL+S+S+S detection frame is 0.838.Compared with the original method,the m AP of new method is improved by 0.012 m AP,which proves that the new method has higher precision and recall.In the difficult object detection,as the detection of overlapping targets and small targets,COL+S+S+S algorithm has achieved better detection results on these specific samples than Cascade?RCNN.At the same time,the detection speed of COL+S+S+S and CL+S+S+S is 10.8 frames per second as same as the original Casacade?RCNN model.In terms of memory consumption,the COL+S+S+S method uses 10452 MB memory and Casacade?RCNN uses 10528 MB memory,which proves that the COL+S+S+S detection frame improves the detection effect without increasing the resource consumption and reducing the detection speed,which makes the new model better than the original model in all aspects.The COL+S+S+S has better application value than the original Cascade?RCNN model.On the public list of Pascal?VOC datasets,the m AP obtained by our algorithm is equivalent to the third place in the list trained only by Pascal?VOC datasets.