Research On Object Detection In Complex Scenes

In recent years,with the rapid development of information technology,enormous data of images and videos are generated on the Internet every day.The analysis and pro-cessing of these explosive increasing data is related to the security of cyberspace and the improvement of user experience.Object detection is one of the basic technologies of computer vision and multimedia applications.It has a very wide range of application requirements in practical engineering,and has a very important research significance.The task of object detection is to recognize all the categories of objects in a given im-age,regress their coordinates of bounding boxes,and use an external rectangular box to locate the position of the identified object.Since objects in the given image may have different categories,scales and occlusions with surrounding objects,it is more difficult and challenging to detect multiple objects in one image than general image classification task.In recent years,with the development of deep learning,object detection has made a breakthrough,but there are still some difficulties in the face of small object detection in complex scenes,detection with very few training data and so on.In this thesis,we have carried out in-depth research on the above problems.The specific research con-tents include:studying the small object detection problem of adaptive convolution in complex scenes from the aspect of dynamic feature extraction;studying object detec-tion of the context information extraction problem from the aspect of sparse attention fusion;studying the few-shot object detection problem from the aspect of dynamic gen-eration of prototype features and auxiliary detection module.The main contributions are as follows:(1)Object detection based on adaptive convolution:An adaptive convolution block is proposed,which dynamically adjusts the param-eters of the convolution kernels according to the input feature map,and then convolves the input feature map with the obtained dynamic convolution kernels.Since the learned convolution parameters depend on the input data,it can selectively extract the most suitable features for the current scene for object detection.After adaptive convolution enhancement,the feature map can pay more attention to the key objects,suppress the interference information of irrelevant objects in the environment,and effectively im-prove the detection accuracy.The proposed adaptive convolution block is lightweight and fast.By integrating the adaptive convolution block directly into the existing de-tection framework,this thesis constructs a new real-time adaptive convolution object detector,which achieves a better balance between speed and detection accuracy.(2)Sparse attention block for object detection:The existing non-local modules model the long-range dependence of image fea-tures by introducing self-attention mechanism.However,due to the need to generate dense attention map,non-local module needs a huge mount of GPU memory and com-putation overhead,which is limited in practical applications.In this thesis,a sparse attention block is proposed.By sampling the most representative features,a sparse attention map is established and context information is aggregated,which can greatly reduce the amount of computation and memory overhead.After searching the local peak response in the heat map of the input feature map,a set of sparse positions are dynamically selected to extract the key elements and value elements,and the sparse dependency relationship between the query elements and the key elements is estab-lished.This sparse attention block can be easily inserted into existing object detection frameworks,and the detection accuracy can be improved with negligible computation cost.(3)Few-shot object detection with dynamic prototype feature generation:Traditional object detection methods usually rely on a large number of training data,but the cost of preparing large-scale high-quality fine labeled training data is very high.In the real-world business,there are often only a few support samples are provided for the new class object detection,which requires a high effecient detection algorithms for few samples learning.This thesis proposes a dynamic prototype feature generation module,which calculates the correlation coefficient between the features of the query image and the features of the support samples,aggregates the support features accord-ing to the correlation coefficient,dynamically generates the prototype features,uses the similarity between the small sample support set and the query set to detect new targets,and suppresses the error detection results in the background.At the same time,the aux-iliary detection module and the source-target domain reciprocating training strategy are proposed to improve the information utilization of a small number of support samples and reduce the risk of over fitting of the detection network in the target domain with a small number of samples.