Small Object Detection Based On Deep Learning

In recent years,the algorithms of object detection have changed rapidly.However,due to the inherent characteristics of small-size targets(such as small physical size,blurred edge information,etc.),small object detection accuracy is very low,and what's more it is difficult to improve,which has gradually become one of the difficult tasks in object detection.In reality,the existence of small objects is universal and more practical.Therefore,it is very important to improve the detection accuracy of small object detection.This article uses the public dataset COCO,which has a large number of small targets.In order to compare the two basic methods for small object detection(based on anchor generation and key point regression),this article uses these three basic methods to develop a small object detection research:(1)Small object detection based on anchor generation and multi-scale adaptive learning.The basic detection method based on anchor box generation is one of the commonly used small object detection methods,but it is usually in the form of fixed anchor boxes.To this end,this paper improves the original method,and proposes an adaptive method of generating anchor frames,which guides the network to use the learning of different depth feature maps to match anchor frames suitable for the target size.In addition,due to the current general data set,pictures containing small objects take up less than targets of other sizes.In order to balance the disparity in the number of small size objects and other size objects,on the one hand,this paper proposes a new balance loss function to punish other size objects during the training phase to increase the network's attention to small object feature learning,on the other hand enhance the multiscale feature learning module,strengthen the network's association with multi-scale feature information,and the AP value of small object detection on the COCO data set can be increased by up to 18%.(2)Small object detection based on key point regression and context sensitive.Considering the shortcomings of the method based on anchor generation,this paper uses the basic detection method based on key point regression,on the one hand,it can avoid these problems,and on the other hand,it also simplifies the series of pre-processing operations required to do early prediction.For the key point regression method,this paper improves the stack hourglass feature extraction module to enhance the basic network's learning of context information around objects.In order to reduce the loss of small object information in the process of network pooling,the direction pooling operation is introduced to replace the original downsampling operation,so that the detection of small size objects has been significantly improved on the COCO data set.(3)Small object detection based on single keypoint regression and scale perception.The commonly used method based on key point regression is generally to predict multiple key points on the image.In this paper,the basic method based on single key point regression is used to simplify the post-processing operation of the network,and the detection speed is improved in the detection result process.In order to further increase the receptive field range of pixels at different depths of the network from the basic structure of the convolutional neural network,a new RE convolution operation is designed to replace some traditional convolution operations for small object detection.The method has higher detection accuracy.