Research On Key Technologies Of Low Resolution And Unconventional Object Detection

As one of the most important research directions in the field of computer vision,a large amount of object detection algorithm research has been proposed.However,when the deep-learning-based object detection algorithm is applied to actual video images,the detection effect is often not ideal.The fact that cases this situation can be divided into exogenous and endogenous reasons.The exogenous reason is that the detection accuracy of object detection models will be greatly reduced when encountering low resolution images,while the endogenous reason is the defects of different object detection algorithms,which leads to a sharp decline in the detection accuracy of the model on unconventional types of detection objects such as extreme sizes and similar densely distributed targets.And these problems are becoming obstacles to promote the application of object detection model.To solve the above problems,this dissertation mainly carries out the following work:(1)To solve the problem that object detection model's performance severely suffers from low-resolution images,this dissertation provides a solution that can organically combine the object detection model with a super-resolution algorithm.This solution firstly designs a back-projection super-resolution network based on dynamically weight prediction algorithm,by which means the network can enlarge and reduce the size of images at any magnification.Then,by establishing information communication channel between the middle layers of super-resolution and object detection structure,the superresolution network becomes a part of the object detection model to improve object detection effects.The experimental results show that the detection accuracy of the object detection model is reduced by 8% and 24% respectively on resolution degraded video images,and the reduction is reduced to 3% and 9% after combining with the superresolution algorithm.(2)Based on the previous work,this dissertation studies how to improve the detection accuracy of lightweight object detection model on low resolution images without adding additional network structure.The part of work proposes a multi-level knowledge distillation approach.The main idea of this method is to train a teacher and a student network on high-and low-resolution images,respectively.And the middle layer of student network is supervised by multi-middle layer of teacher network,so that the student model can mimic how the teacher network works.At the same time,this method designs a network structure combining with pooling layer and fully connected layer to judge which part of the information transferred by teacher network is important and which part is redundant.By doing so,the knowledge transferring operation can be much more efficient.The experimental results show that the accuracy reduction of lightweight object detection model on resolution degraded(4x and 8x)video images is reduced from 14% and 49% to 8% and 39% by training it with multi-scale knowledge distillation algorithm proposed in this part of work.(3)After improves the detection accuracy of object detection model on low resolution images,it is found that the detection accuracy of object detection model realized by anchor-based path decreases significantly when encountering objects with abnormal sizes that often appear in the actual situation.Therefore,an automatically anchor learning algorithm that can dynamically predict anchor boxes instead of using a fixed anchor configuration is proposed.The algorithm realizes the object detection task by dynamically predicting an anchor of any size for each pixel position on the feature map.The overall process of the algorithm can be divided into three steps: position prediction,feature degradation and anchor prediction.During detection,only one anchor box is predicted for a limited number of pixel positions to complete the object detection process,instead of placing anchors for each position on the feature map like the ordinary anchor-based model,which greatly reduces the computational burden.The experimental results show that due to the uncertainty of the predicted anchor size,this method can fit to any size target,and effectively improves the detection accuracy of the model by around 3%.(4)In addition to the problem of anchor-based model,some object detection models that use anchor-free path also show poor detection effect on densely arranged similar targets,which can be frequently observed in practical application.Therefore,this dissertation proposes a key point matching algorithm based on the division of prediction center position.Different from some anchor-free based models that use appearance information to match key points,this method completes the tasks of key point matching and object detection by fully considering the spatial information between the box composed of predicted key points and the ground-truth target box.In the process of detection,some prediceted boxes can find the real target boxes that are close to their center points.And a loss function is designed for this kind of boxes,so that the positioning results can be obtained by regression on them.And the secondary classification can be carried out to obtain more accurate classification results.The other part of the predicted box is far away from center points of any ground-truth boxes.Therefore,a loss function is designed to minimize the number of predicted boxes in this part,and the predicted boxes should also be far away from other types of predicted boxes as far as possible,so as not to confuse the network model.Experiments show that using the key point matching method based on the division of prediction center position can improve the detection accuracy of the model by about 5%.