Research On Object Detection And Application Based On Weakly Supervised Learning

An important branch of the computer vision field is object detection.In the emerging application fields of artificial intelligence represented by intelligent surveillance and unmanned driving,the requirement for object detection systems is to achieve boundary extraction,classification and recognition of various objects in real images.Wait.With the rapid development of convolutional neural networks,the research of object detection based on deep learning has become an important means to achieve high precision and high accuracy of detection.At present,common object detection algorithms usually require pixel-level annotation information to train network models,but the acquisition of pixel-level annotation information is labor-intensive and easy to be restricted by conditions so that object features are submerged in noise,which is not universal.Therefore,object detection algorithms under weakly supervised learning are gradually being favored by researchers.Through investigating the object detection algorithm under weak supervision learning,it is found that the detection accuracy of the existing weak supervision detection algorithm is far from that of the full supervision algorithm,and the detection rate is too low to complete the real-time detection task.In view of the above problems,the main work of thisis as follows:(1)Research on the weakly supervised object detection model based on pixel gradient map.Because the object detection model under weakly supervised learning lacks the label of location information during the training process,the accuracy of model detection is low.This article first uses the different response of the image to each category to generate the pixel gradient map of the image for each category to obtain the approximate position of the object in the image.Second,design an iterative similarity mining method to obtain more complete location information.Finally,use the obtained position information to generate a bounding box and pass it into the fully-supervised network model.Solve the problem of the lack of labeling of location information in the training process,and improve the accuracy and detection rate of object detection under weakly supervised learning.(2)Research on the positioning optimization of the weakly supervised object detection model based on the pixel gradient map.Because the bounding box passed into the fullysupervised network cannot completely coincide with the manually labeled bounding box,the detection result of the model often contains only part,and complete detection information cannot be obtained.Thisis introduces Io U-Net to further optimize the positioning of the model,and first learn the predicted Io U network.Then use Io U-Net to guide the generation of NMS to optimize the generation of bounding boxes.Finally,the fine pooling layer is used to replace the pooling layer in the fully-supervised network for joint training to optimize the positioning performance.Solve the problem of inaccurate detection of the bounding box during the training process,and avoid the suppression of the precise bounding box.Compared with algorithms that do not use positioning optimization,the accuracy of object detection under weakly supervised learning is improved,but the increase in network models reduces the detection rate.