Research On Object Detection Based On Convolutional Neural Network

Object detection,as one of the most important tasks in computer vision,aims to recognize and locate specific objects from static images or videos.The realization of the object detection integrates image processing,machine learning,deep learning and artificial intelligence technology.It is widely used in military drills,medical image analysis,intelligent transportation,industrial testing and human-computer interaction and so on.However,there are both opportunities and challenges in the application of object detection due to the variety of object shapes,occlusion and complex environment.In this paper,two kinds of object detection algorithms are researched,namely object detection based on region proposals and object detection based on regression.This paper proposes cascaded convolutional neural network for object detection to solve the problem of large localization bias and large portion of background samples in object detection methods based on region proposals.Proposed RefineNet can effectively reduce the localization bias and increase the detection confidence by iterative localization and a cascaded classifier can effectively decrease the false positives by integrate N binary classifier to make the network focus more on intra-class variance.A multi-feature fully convolutional network is proposed aiming to solve the problem of low precision in small objects and poor localization.Framing object detection as a regression problem can simplify detection pipeline and improve the detection speed.Multi-feature concatenation can efficiently fuse shallow and deep information and increase the detection confidence.In order to validate the effectiveness of proposed methods,PASCAL VOC database are used to perform the object detection experiment.In this paper,the precision of each class and the average precision of all classes are measured.The experimental results show that the proposed cascaded convolutional neural network outperforms the Faster R-CNN framework by 3.1%,while the multi-feature fully convolutional network outperforms YOLO by 10%,proving the effectiveness of the proposed methods.