Research On Target Detection Algorithm Based On Sparse Learnable Proposal Regions

Target detection is to identify the interested objects in the picture,this is a crucial technology for many complex scenes of computer vision.Different scenes need different frameworks of target detection algorithms,but all of them have put forward their own requirements on the performances of accuracy and real-time,such as unmanned driving,security system etc.In recent years,there have been born a variety of target detection algorithms based on convolutional neural network,however,the target detection algorithm with high accuracy has lots of computational steps because of its complex design,so it needs a lot of time,and some high real-time algorithms have a poor accuracy by short steps.The traditional algorithms generate a large number of proposal boxes on the picture at the first step,they usually have a fixed size,then the algorithm will go on for the further testing which based on these boxes,so we need to set up many parameters for the model in advance.Besides,these boxes will bring huge imbalances between the positive and negative samples,so we need to use the way of non-maximum suppression,to filter out some redundant boxes.Therefore,the most important thing is to design a target detection algorithm which both has high accuracy and real-time in the academia and industry.In order to simplify the flows and improve the learnability of the model,based on traditional ways of producing the alternative boxes,a target detection algorithm based on sparse learnable proposal regions is proposed.It uses a feature pyramid network to optimize the quality of the original image firstly,after this network,the image will have more messages,but it also keeps the original size.Based on the output image,the algorithm directly generates a lot of sparse proposal boxes,then sets automatic learnable parameters for each proposal box,so that the model does not need any other processing step,the proposal box is no longer unlearnable and fixed.Aimed to be more accurate,the optimization module is added into the model,in the module,parameters of the sparse proposal boxes can be adjusted,according to the specific features.Later work will be carried out on these sparse learnable proposal regions,which has improved the accuracy and the speed of detection and recognition.In order to test the feasibility of the algorithm,by working on COCO and VOC dataset,and the pictures taken in the school,we compare this algorithm with the traditional algorithm,it proofs that this algorithm is better than other traditional algorithms by the experiment.Removing module?the number of original proposal boxes and the time of iteration was also discussed.In addition,some experiments are carried out on the embedded platform RK1808,by building the development environment and algorithm writing,we put the designed network to an embedded system,and input some pictures of different environments,the more accurate results have been obtained,which proved the effectiveness of the proposed method.