Research On Target Detection And Recognition In Motion Blurred Images Based On Target Region Proposal

Object detection has always been a very challenging research topic in the computer field and a very important branch of computer vision.It has a wide range of applications in today’s daily life,criminal detection,military object detection,and medical operation.It also has critical research value for the future unmanned era.With the continuous development of computer technology,there is a higher requirement for the detection and recognition accuracy.In practical applications,especially in complex scenes with poor image quality,the positioning and recognition of multiple objects also faces many challenges.The traditional object detection algorithm needs to manually obtain the object feature information from the original image,and the algorithm is complex,time-consuming and inaccurate,which leads to the research of object detection has been stagnant.In recent years,with the development of artificial intelligence,the object detection algorithm based on deep learning has achieved high detection accuracy,but it still struggles in small-size object detection.Motion blur is a common phenomenon of image degradation in daily life.How to detect and recognize motion blur images caused by camera shake,object motion or out-of-focus has become a realistic problem to be solved.These problems need further exploration and research.At present,the deep learning-based object detection algorithms use the idea of object proposal to focus on the positions of the image that most likely contains objects,greatly reducing the time complexity and computational complexity of the algorithm.At the same time,for the research of blind motion deblurring,the method of deep learning has also achieved the latest results.Based on the above research,this thesis proposed a motion blur image object detection and recognition model based on the object proposal for the existing problems.The specific work is as follows:(1)For the problem of small-scale object detection accuracy,this thesis proposed a multi-scale and multi-task implementation method for object detection.According to the characteristics of the deep and shallow features of the convolutional neural network,and the respective requirements of the two tasks of classification and detection,the combination of deep features and multi-scale features can effectively exploit the advantages of different feature information.In addition,the proposed multi-scale fusion method has fewer parameters and low design complexity.It has very good results on the PASCAL VOC dataset,especially in the detection of small-size objects.(2)For the deblurring task,a new Scale-recurrent Network is proposed.In the model construction,we use the encoder-decoder and fewer residual connection blocks as the intermediate layer,which greatly reduces the network scale.It has the advantages of less parameters and short training time.In addition,on the GOPRO test dataset,the proposed model is quite competitive in terms of image restoration quality compared to current state-of-the-art methods.In terms of speed,the single picture detection speed is 0.17s,which greatly exceeds the existing works.(3)An object detection model for motion blurred images is proposed.Deblurring and object detection are effectively combined for joint optimization using a 7-step experimental procedure.For the verification of the proposed algorithm,due to the lack of motion blurred image datasets that can be used for object detection,this thesis uses blurry images generated by convolving synthetic blur kernels on the object detection datasets,then obtains a dataset for testing,and finally quantitatively analyzes the proposed network.The results show that the combined network can effectively detect and identify motion blurred images.