Research On Object Detection Algorithm And Application Based On Deep Reinforcement Learning

In recent years,with the support of deep learning technology,object detection has become the focus of the development of perceptual intelligence in the field of computer vision,and has been widely used in many fields,and has been listed as a research hotspot by academia and industry.The development of target detection has important social significance in people’s lives,and the intelligent applications derived from it as a core basis have an important impact on people’s life patterns.As an important cornerstone of vision in perceptual intelligence,its problems and challenges in the process of research and application have become the key topics of academic research.At the same time,it has also shifted from confirmatory research in academia to practical applications in industry.To summarize and summarize the existing research work of target detection,it can be found that there are many related researches on target detection in training sample label allocation,target recognition in special scenes,and object expression.Although object detection has been developed for many years as a basic research topic in the field of perceptual intelligent vision,it still faces many problems and challenges in the process of deepening the practical application in the industry.In the work of this article,we analyzed several key issues of current target detection,summarized the development path of target detection and its current research status;and on this basis,we focused on the division of training samples,the learning of complex samples,and the formal expression of samples.These three key issues are analyzed and researched and corresponding solutions are given.In the work of this article,it not only includes theoretical analysis and derivation research,but also includes deep analysis and mining of method and practice.The main contributions of this article include the following three aspects:1.Training sample division: For the problems caused by the imbalance of positive and negative sample categories in the target detection method,the strategy of training sample division as a research direction has a significant impact on the performance of the target detection model.The reasonable division of training samples can make the target detection model.Focus more on the target instance itself in the image to improve model performance.As the foreground of the detected target instance,compared to the background occupying a lower proportion of the space in the image,dense prediction is the main method used in current target detection to generate a large number of negative training samples,resulting in a more simple model training effect Negative samples fail to detect the target instance that should be paid attention to.This problem is more serious in professional applications such as small-scale medical image background complex object instances.In order to solve this problem,we propose a training sample division strategy RTSA based on deep reinforcement learning control.With the support of this strategy,the threshold for dividing the positive and negative training samples will be dynamically generated and determined according to the statistical attributes of the generated anchor frame set,and can be optimized by deforming the anchor frame of the negative training sample by the agent under the control of deep reinforcement learning Increase the proportion of positive training samples,so as to solve the problems caused by the imbalance of positive and negative samples in a targeted manner.Our proposed method was verified on the laparoscopic medical data set,and the ablation experiment further verified the influence of different settings of the method on the effectiveness of the performance of the method.2.Complex sample learning: Rotating target detection faces the problem of excessive background interference in images with high background information.Therefore,as a research direction,complex sample learning has a significant impact on the performance of target detection models.When horizontal target detection detects targets of any orientation,there is too much redundant information in the bounding box,which causes the network training of the model to require more training data and time to improve the robustness to interference information.In order to solve this problem,we propose a target detection method ORL-RPN based on deep reinforcement learning control optimization.The target detection task is converted from a usual regression problem to a sequence decision task that controls a series of decomposed image detection actions.Learning and polar coordinates of the target expression form,the agent can use dynamic strategies to accurately sample the rotating target.In order to verify the impact of sample complexity on the target detection model,we adopted a dedicated rotating target detection data set,and conducted comparative experiments on these data sets.Through experiments,we verified the impact of sample complexity brought about by rotating targets on target detection,and confirmed the effectiveness of our proposed target detection method ORL-RPN.3.Formal expression of samples: The expression of object instances in images is a key research topic in target detection tasks,and has always been concerned by people in the academic field.The expression of the sample,that is,the definition of the expression of the object instance,has an important influence on the target detection because it determines the design and realization of the target detection method on the model.Affected by traditional methods of target detection,box expression has always occupied a dominant position in sample expression.People use the frame as a reference,and use regression methods to predict the target instance to achieve the purpose of including the object in the bounding box.However,with the development of the target detection field,the sample expression of this bounding box has become the bottleneck of the performance improvement of the target detection method.The bounding box can only express the classification information and rough position information of the objects in the box,and the expression of the box Because of the restriction on the expression of object shape,the difficulty of image location search is increased.In order to solve this problem,current academic circles have developed point-based expressions to replace the original box-based expressions.However,the points expressing the target instance still face sub-optimal problems in the distribution and quantity.For this reason,we propose a target detection model RLRep Points based on deep reinforcement learning to control the expression of key points.Through experiments,we verify that under the optimization of reinforcement learning,the points expressed for the sample are adaptive to the optimal state and the robustness of the target detection model is enhanced.As an important part of the target detection method,sample learning plays a key role in the performance of the model.This paper analyses the impact of sample learning on target detection models in terms of training sample partitioning,complex samples,and sample form representation,and improves the performance of the model by incorporating a deep reinforcement learning module.