Research On Deep Learning Based Object Detection Algorithm

Object detection is an important research direction in the field of computer vision.Its application is essential in many fields such as intelligent video surveillance,driverless technology,and computer-aided diagnostic technology.However,due to the influence of many factors such as deformation,occlusion,angle and environmental changes in the actual scene,object detection is still a challenging task.And how to design a feature which can accurately identify the object without being influenced by various external factors has become the focus of research in this field.Traditional object detection algorithms often use artificially designed feature algorithm to extract features.However,these extracted features contain relatively shallow information and have poor performance in detection performance.Moreover,designing feature extraction algorithm often requires quite a lot background knowledge and solid professional knowledge,which makes it difficult to design a good feature.The emergence and development of deep learning has solved this problem well.The deep learning algorithm can automatically learn from a large number of samples,process and combine the extracted features for different tasks to learn representations of data with multiple levels of abstraction.Based on the thorough investigation of the related technologies of deep learning and object detection,this thesis analyzes the existing problems and improves the existing deep learning based object detection algorithms.The research contents of this thesis include the following aspects:(1)Based on the research on the existing deep learning based object detection algorithm,it is pointed out that most of the existing algorithms use the same feature map to predict location detection and category.Then we analyze and point out that the location prediction task and category prediction task in object detection problem have different requirements for feature characteristics.To solve this problem,based on the object detection algorithm SSD300(Single Shot MultiBox Detector),we redesigned the network structure is and named it DF-SSD(Decoupled Fused Feature Based Single Shot MultiBox Detector).The new algorithm separates the location prediction and class prediction feature to a certain extent.Considering that location prediction requires more details information,DF-SSD also merge low-level feature maps from the dilated convolution into the feature maps used in location detection.(2)Since the features used in SSD300 algorithm are deep semantic features and small object detection needs more shallow displacement sensitive features,SSD has low detection accuracy for smaller objects.Therefore,it is necessary to increase the number of small object in the training set by means of data augmentation methods.The method used by SSD300 introduces noise and wastes a lot of storage space.Aiming at this problem,this thesis proposes a method to increase the number of small object samples based on pyramid image sequence which can effectively increase the number of small object samples without introducing any noise.In the new data augmentation algorithm,first,the original image is sampled at two different ratios to form the image pyramid sequence,then the two level pyramid image is spliced together to ensure the same size as the original image.(3)Based on the above research,a simple object detection system is realized so that the detection effect can be intuitively perceived and the performance of quantitative measurement can be obtained.Multiple control experiments were designed based on this system.Experimental results show that the proposed separation features,feature fusion,and pyramid image sequence data augmentation all improve model performance to different degrees.The DF-SSD achieved an mAP of 78.7% on the Pascal VOC 2007 test set,with an increase of 1.5% on the basis of the SSD300.