Object Detection Based On Deep Learning

All along,image recognition is an important field of pattern recognition,and the application of image recognition has also penetrated into all aspects of life,such as medical image analysis used in the field of health care,such as pedestrian recognition used in video structuring,and so on.How to locate and recognize objects in an image is one of the basic problems in the field of image recognition and one of the underlying technologies in the field of computer vision.In the actual industrial production applications,in order to ensure the feedforward speed of the detection network,one-stage object detection method is often used.However,the accuracy of one-stage method is often inferior to that of two-stage method.In order to further improve the accuracy of one-stage object detector SSD(Single Shot MultiBox Detector),this paper mainly improves the performance of SSD in two aspects.The first point is that common deep learning-based object detection methods usually introduce loss functions at the end of the network to guide network training.In order to further improve the detection performance,this paper also introduces additional monitoring information,and uses the loss function based on image segmentation to assist network training,thereby optimizing the training direction of the network and improving the accuracy of the detector.The second point is that the low level of convolutional neural network is rich in shallow geometrical details,but less semantic information,while the high level of convolutional neural network is rich in semantic information,but lack of geometric detail information.In the SSD structure,shallow features are used to detect small objects.Due to the lack of semantic information,there are many missed detections on small object targets.By designing the multi-scale semantic feature fusion structure,this paper adds semantic information to the shallow features,and uses the attention mechanism in the prediction stage to improve the feature extraction ability and effectively improve the object detection performance for small obj ects.This paper further explores the application of the proposed method in pedestrian detection,discusses and analyses the problems of the loss function Repulsion Loss used to solve the occlusion problem in crowd pedestrian datasets,and proposes corresponding solutions.Finally,this paper designs and implements a demo system of object detection for crowd counting,which shows the performance of the proposed method intuitively.