Research On Models Of Three Granularity Image Recognition Based On Convolutional Neural Networks

Image recognition is a very important research task of computer vision.The use of computers to complete the recognition of targets or objects in different modes has been a valuable part of artificial intelligence.The image recognition models refer to training and verification to produce a recognition model using images and the annotation information,which should have a relatively robustness and mobility.According to the granularity of pre-recognized targets,image recognition tasks can be divided into image-level target recognition,regional-level target recognition and pixel-level target recognition.Therefore,this study focused on the above three recognition categories and conducted research from the following three aspects:1.Image-level target recognition task.On the basis of the VGG model,this paper proposed a HRST-Net model for high-resolution small target recognition,which can effectively reflect the location information of small targets while improving the classification accuracy.The existing models applied to high-resolution small target image lose the key features of the target and cannot display the target location information effectively in the image zooming.The improved Oxford Visual Geometry Group network in the proposed model can effectively solve the problem of target feature missing in the process of image zooming when the size of the target is far smaller than the image size.At the same time,the improved Grad-CAM key area recognition algorithm can complete the visual display of the target in the form of thermal map.In this paper,the effectiveness of the model is verified by an experiment of pulmonary tuberculosis recognition on chest X-ray images.While ensuring good performance,the recognition accuracy reaches 83.5%.At the same time,this paper also uses CT images in the identification of lung inflammation for model verification,and the final accuracy has reached 78.2%.2.Regional level target recognition task.This paper uses the lung nodule recognition experiment of chest X-ray images to analyze the problem about the dependence of recognition models such as Faster RCNN and YOLO on data annotation.In order to solve the problem of over dependence of existing recognition models on data annotation,a processing model of incomplete annotation data set was proposed.The model can use the incomplete labeled dataset to generate the complete labeled dataset,so as to improve the recognition accuracy of regional targets.The existing target detection models require that the training dataset must contain complete manual annotation information,otherwise the target features will be missing or confused.The proposed model can effectively solve this problem and fully retain the key features of the target.In this paper,we took the pathological image data set with incomplete annotation as an example,experiments show that the datasets with 70%and 50%of the labeled information are preprocessed by unCL-GAN model,and then learned from the existing regional recognition model.The recognition accuracy of the obtained model has been significantly improved,which makes the final recognition accuracy rate reached 80.9%and 75.7%.The recognition accuracy rates are close to completed annotations.3.Pixel-level target recognition task.Based on the unCL-GAN model,this paper proposed an improved model named NNI-Net,which can more effectively complete the task of target detail feature recognition and improve the accuracy.This task requires that the recognition accuracy be standardized to the pixel level accuracy,so it is necessary to improve the existing model to improve the accuracy of the model for pixel level target recognition.In this paper,the U-Net structure is integrated into the generative countermeasure network,the loss of pixel decision is modified,and the user-defined upper sampling layer is added to capture the detailed features,so as to improve the recognition accuracy of the model.In this paper,we take the retinal vascular recognition of fundus images as an example,and compare it with the existing models.The results prove that the model proposed in this paper can achieve the best recognition accuracy of 0.963 and 0.973,respectively.At the same time,the lung contour recognition of the chest X-ray image was used to verify the model again,and the final recognition Dice coefficient reached 0.972±0.005.Based on the image recognition models,this paper proposed models to improve the performance in three different granularity target recognition tasks:image level,region level,and pixel level.Among them,the highresolution small target recognition in the image-level target recognition provides ideas for the construction of the regional-level target recognition model,and thus expands the target recognition model based on incomplete annotations.The model includes two branch structures.A branch structure can support the research of pixel-level target recognition models.Experiments proved that the proposed models can effectively solve the problems encountered in data preprocessing,model building,and prediction accuracy in other application scenarios,and provide new ideas for the image recognition models based on convolutional neural network.