A Study On Multi-scale Feature Extraction And Optimized Non-maximum Suppression For Object Detection

As a basic and challenging task in the field of computer vision,object detection has important academic value and practical significance.Its task is not only to understand the semantic content of the image,but also to accurately find the specific location of the target in the image.Traditional methods usually use hand-designed SIFT(Scale Invariant Feature Transform),HOG(Histogram of Oriented Gradient)and other features to describe images,and use classifiers to classify them.With the development of deep learning and the superiority of convolutional neural networks in processing image data,the field of target detection in recent years has entered the era of deep learning.This thsis studies the object detection based on deep learning.Aiming at the problem of insufficient multi-scale information in the feature extraction stage and the lack of location confidence in the non-maximum suppression algorithm in the post-processing stage,the object detector RFF-Retina Net(Receptive Field Fusion Retina Net)is proposed,which can effectively improve the accuracy of target detection.The main contributions of this article are as follows:(1)Aiming at the problem of multi-scale feature extraction,a multi-scale feature extractor based on the receptive field fusion module is proposed.The extractor consists of three parts: the backbone network,the receptive field fusion module and the feature pyramid.First,the backbone network extracts the basic image convolution features,and then the receptive field fusion module enriches the multi-scale information of the features,and finally uses the feature pyramid to fuse the multi-scale information of different layers to obtain multi-scale features.(2)Aiming at the problem of the lack of confidence in the fixed position in the non-maximum suppression algorithm,an optimized non-maximum suppression algorithm based on joint confidence ranking is proposed.Different from the traditional non-maximum suppression algorithm that uses the classification confidence of the classification subnet output as the sorting criterion of the bounding box,we consider the equal importance of the classification confidence and the fixed position confidence,and use the output of the classification subnet and the positioning subnet The joint confidence of the bounding box is designed for the output of,and the joint confidence is used as the sorting criterion in the non-maximum suppression algorithm.(3)Based on the above two innovative modules,the object detector RFF-Retina Net is proposed.RFF-Retina Net uses a multi-scale feature extractor based on the receptive field fusion module to fully extract the multi-scale features of the input image.Then,the multi-scale features of different layers are respectively sent to the classification subnet and the location subnet for classification prediction and location offset prediction.Finally,using the output of the classification subnet and the positioning subnet,the classification confidence and the fixed location confidence are respectively calculated,and input into the non-maximum suppression algorithm based on the joint confidence ranking to obtain the final detection result of the model.(4)Experiment with RFF-Retina Net on the object detection data sets PASCAL VOC and MS COCO,and compare it with two-stage detectors such as Faster R-CNN and one-stage detectors such as Retina Net that use a variety of backbone networks.Among them,RFF-Retina Net achieved the best detection effect when using ResNeXt-101 as the backbone network,and the m AP value reached 42.3.At the same time,we also designed a series of ablation experiments to verify the effectiveness of the two innovative modules.