| Object detection is an important branch of computer vision. Its main goal is to detect and locate specific targets from static image sor videos. It integrates image processing, machine learning, pattern recognition and artificial intelligence, and is widely used in intelligent transportation, medical image analysis, hum an-computer interacti on and other fields. Because of the diversity of target environment and the variety of target shape, occlusion and so on, there are both opportunities and challenges in the application of object detection.Aiming at resolving the shortcomings of using traditional manu al feature extraction method to get low accuracy of object detection,a new target detection algorithm based on convolutional neural net work is proposed on the basis of deep learning in this paper. In or der to solve the problem that the input image frame of the convolution neural network has low accuracy and can not acquire the more essential features of images, this paper divides the algorithm into three steps: (1) extracting the region of the image which may contain the object as the candidate window by using the multi-scale fusion method of multi-level segmentation; (2) optimizing the object proposals using the position deviation of the candidate window, which is determined by the intersection of the super-pixel and the candidate window; (3) in order to adapt to the input size specified by the convolutional neural network, this paper proposes a new method of partially filling the candidate window by using the super-pixel across the candidate window, this method further improves the accuracy of the extracted feature of the network model; (4) training SVM linear classifiers for each object class on the basis of the extracted features. By improving the accuracy of input of the network model,the accuracy of feature extraction is improved, and the feature is extracted by using the fine-tuning convolutional neural network model, which improves the accuracy of the detection result compared with the traditional manual extraction feature.In order to prove the effectiveness of this algorithm, we use PASCAL VOC 2007 database to perform the object detection experiment. The database provides 20 kinds of common objects, which are divided into training verification and test images. It can verify the effectiveness of the algorithm. In this paper, the precision of each category and the average precision of all classes are calculated. Our experimental results show that the proposed algorithm outperforms the R-CNN framework by 3.4%, which proves the effectiveness of the proposed algorithm. |
