Research On Small Target Detection Method Based On Deep Learning

Object detection is the basic topic in the field of computer vision,it is the cornerstone of pedestrian detection,vehicle detection,and instance segmentation,and has important research significance.The main task of object detection is to locate the position of the object in the input image and the category of the object.Through the analysis of the current domestic and foreign research results,we discuss the main challenges and deficiencies of object detection.In order to solve the challenges and deficiencies,this topic is based on the classical algorithm Faster R-CNN in deep learning.Starting from image preprocessing,feature fusion and convolutional neural network for feature perception area,we propose an effective object detection algorithm to solve the problem of low accuracy of small object detection and improve the detection effect.We propose an HF-FF R-CNN algorithm based on multilayer frequency domain feature fusion.The aim is to solve the problem of poor small object detection effect caused by poor image quality,small target detection difficulty,large change in target scale,and loss of image detail texture in the process of feature extraction.The HF-FFR-CNN algorithm adds high frequency enhanced image(High frequency enhanced image,HF)and feature fusion(Feature fusion,FF)modules to the Faster R-CNN model.In the image pre-processing stage,the high-frequency information of the target is added to the image,and the high-frequency enhanced image and the contrast-enhanced image are used as the input samples of the algorithm to enhance the feature expression ability of the target in the training data.For targets with smaller pixels,we change the anchor scale in the RPN,so that the anchor frame generated by the RPN is more consistent with the characteristics of targets in the image.Meanwhile,feature fusion modules are introduced in the backbone network section to obtain low-resolution semantic information features and high-resolution detail features,and it solves the problem of disappearing features of small objects in the backbone network,and further improves the detection effect of small objects.Experimental results show that the proposed HF-FF R-CNN algorithm has good performance on DAGM 2007 dataset of ten kinds of small targets,with an average detection accuracy of 97.9%,and is significantly better than the original Faster R-CNN algorithm in PASCAL VOC 2007 test dataset.Meanwhile,we propose a region-sensitive network structure that decouples the classification and localization tasks based on the HF-FF R-CNN algorithm,with the aim to solve the mapping bias caused by multiple quantification of the RoI pooling layer,the different sensitivity of the image classification task and the localization task to the object features.The image classification task is more sensitive to the place where the object is most discriminative,which requires the CNN to have translation invariance,while the positioning task is more sensitive to the object edge,which requires the CNN to have translation variability.Based on this,we first introduce the formable RoI Align module,get corrected features for classification and positioning tasks,and redesign the network branches,classification branches use a fully connected network to ensure that the extracted features have translation invariance,positioning branches use a convolutional layer to ensure more accurate positioning effect.Then four sets of ablation experiments were designed on the DAGM 2007 dataset to verify the contribution of deformable RoI Align modules and branch decoupling to the model.Finally,the two modules are integrated with the proposed HF-FF R-CNN model and compared with the mainstream small object detection algorithms Grid R-CNN,PANet,Cascade R-CNN and YOLO V4 on the PASCAL VOC 2007 dataset to verify the effectiveness of the algorithm for small object detection.