Research On Object Detection Algorithms Based On High-Resolution Images

In recent years,computer vision technology has flourished.As a representative research direction,object detection has played an important role in transportation,medical and military fields.With the improvement of computer hardware capabilities and the continuous innovation of object detection algorithms,the precision and efficiency of detection have also been continuously improved.In object detection,the commonly used datasets are MS-COCO and PASCAL VOC,etc.Most detection methods are proposed based on these datasets,whose common feature is that the image resolution is relatively low.The average resolution of each image is only about 500 × 500 pixels.However,the improvement of hardware technology has also led to an increase in the resolution of captured images.Although these detection algorithms have good performance on low-resolution datasets,they will face considerable challenges on high-resolution images.On the one hand,the real image scene may contain many small objects.These small objects occupy a small aera and are easily missed for detection,which in turn affects the detection precision.On the other hand,the increase in image resolution will lead to a large increase in the amount of calculation,which will affect the detection efficiency.Therefore,detection of small targets and detection efficiency are undoubtedly two major difficulties in high-resolution images.Based on the above difficulties,the main works of this dissertation are as follows.(1)Traditional YOLOv3 uses a multi-scale detection method to integrate the feature information between different feature maps to improve detection precision.However,the three scales are all concatenated to the deep structure of the feature extraction network,resulting in the loss of feature information of some small objects and missed detection in high-resolution image.Therefore,this dissertation adjusts the multi-scale detection part of YOLOv3 upwards,and concatenates two of the scales with the shallower network to retain more information about small objects,so as to improve the detection ability of small objects.Meanwhile,a dilated convolution module is added to the feature extraction network,and three convolution groups with different dilated rates are used to integrate features.For small objects,the spatial context information of the object can be used to further improve the detection precision of small objects.(2)Limited by computer hardware resources,higher-resolution images cannot be directly used as the input of the object detection network.However,If the high-resolution images are down-sampled directly,the target information will also be lost.This dissertation first divides the high-resolution image into several sub-images,and then introduces a coarse-to-fine detection strategy,which greatly improves the efficiency of detection.This strategy adds a coarse detection module to the feature extraction part to filter out some sub-images that do not contain objects.The coarse detection module calculates the confidence that the sub-image contains the objects.For the sub-images with high confidence,it indicates that the objects are more likely to exist,while the sub-image with low confidence is likely to be the background area.In the subsequent feature extraction,Only the sub-images with high confidence level continue to perform feature extraction.Finally,fine detection is performed on these feature maps.By filtering the subimages,the detection efficiency is greatly improved.(3)In order to further improve the efficiency of object detection in high-resolution images,this dissertation proposes an object detection method based on image distillation inspired by knowledge distillation.Different from the traditional knowledge distillation idea,this dissertation uses the sub-images as the input of the teacher model,and downsampled images of the sub-images according to a certain proportion as the input of the student model.The structure of the teacher model and student model are consistent.This method achieves the goal of achieving higher detection efficiency while maintaining higher detection precision through the guidance and training of the teacher model to the student model.This dissertation has 25 figures,13 tables,and 56 references.