Multi-scale Object Detection Algorithm Based on Faster R-CNN

Abstract

Listen

Translate article

AbstractMost object detection methods have low detection accuracy for small and dense objects and susceptibility to noise interference. In response to these problems, this paper designs an improved multi-scale object detection algorithm based on Faster R-CNN, aiming to improve the detection accuracy of small and dense objects. This paper introduces depthwise separable Convolution to reduce the number of network parameters. The amount of calculation uses dilated Convolution to improve the feature extraction of small objects and dense objects by increasing the receptive field, uses Convolution instead of 3 × 3 conventional Convolution to solve the problem that the feature extraction of small objects caused by Convolution is easy to be lost, distorted or the feature extraction redundancy of large objects is too high. The spatial attention mechanism is introduced to efficiently screen features that are more beneficial to object detection and improve the model's ability to deal with large object scale differences in multi-scale object detection. The network model in this paper was trained and verified on the datasets VOC2007 and VOC2012, and the mean average precision reached 86.3%.KeywordsMulti-scale object detectionFaster R-CNNInvolutionSpatial attention mechanism