Research On Small Objects Detection Method Based On Deep Learning

With the development of computer intelligence requirements,target detection as an important part of computer vision has also become a research hotspot in theory and application in recent years.Because small targets widely exist in large-field images,long-distance imaging images,and special-category object images,the research on small target detection has great practical significance in the intelligent location and recognition of image objects.At present,small object detection based on deep learning has higher accuracy and faster speed than other methods when detecting visible light images with certain features,but there is still much room for improvement in network structure methods.Due to the low resolution of small objects,less information,and the weak ability of deep network to express the features of small objects,the research on small object detection is very difficult.In response to the above problems,this article mainly focuses on the research of small object detection methods based on deep learning,focusing on the network structure of small target detection,the training strategy of small target detection,the network framework of small target detection,and the engineering application of small target detection.The work is divided into the following three parts:1.For the two-stage detectors,in order to solve the problem of difficult extraction of small target features at multiple scales,the Faster R-CNN and Mask R-CNN are optimized in structure,and the FPN structure is used to improve the model.Res Net50 and Res Ne Xt101 are respectively used as the backbone of the model to improve the detection accuracy of the network.Besides,we improve the network training method with cross-training,warm-up,momentum adjustment and other methods.2.For the two-stage detectors,to solve the problem of unstable model migration,a layer-wise training method is designed.The improved Faster R-CNN network and the improved Mask R-CNN network are trained layer by layer.We freeze different layers of the pre-trained network and training the remaining network for difficult small targets to improve the detection accuracy and recall rate of the network for small targets.3.For the one-stage detector,to meet the needs of real-time target detection,we train the YOLOv4 tiny network by multiple self-built data sets.The improved network framework and data augmentation are used to compare the accuracy of different target sizes,realize the real-time detection of small objects.Besides we apply the networks on the hardware platform based on TX2.