Research On Image Recognition And Conversion Algorithms Based On Weak Supervised Learning

All kinds of tasks in the field of computer vision are essentially to help people better understand pictures.If we can find the correlation to each other and make use of it,it will promote the completion of each specific task.At present,all kinds of algorithm models in the field of computer vision are not ideal for image processing in natural scenes.Taking Pascal VOC data set as an example,images often have multi-objective,different target scales,complex background and so on.Image recognition and translation is an important research topic in the field of computer vision.The processing effect of related algorithms for natural scene images needs to be improved.At the same time,whether the label information of recognition model can be used to improve the experimental effect of conversion tasks in the field of image translation is also the focus of this paper.Owing to the limitation of training data set,the general object classification model is often ineffective in testing natural scene pictures.Therefore,this kind of model has poor generalization effect and can not adapt to different task requirements.Aiming at the problem that the existing mature object classification model trained on ImageNet dataset can not migrate directly to Pascal VOC dataset,this paper applies VGGNet object classification model to VOC dataset by using the migration learning method combined with the solution idea and model framework of weak supervised object detection task.This paper completes the task of natural image recognition,and designs a comparative experiment to verify the feasibility of the algorithm.At present,image conversion model also has some problems,such as difficult downconversion in natural scenes,image distortion and so on.Image conversion model can not only focus on specific scene objects.For example,in unsupervised circumstances,if the images are not paired or aligned,the network must also understand the specific areas to be converted in the scene.In this paper,the weak supervised recognition model is used to provide approximate position constraints for object transformation in source and target domains,and guide the CycleGAN generator to generate images with less difference from real images.This paper completes the optimization of CycleGAN model,and compares the effect of the image generated before and after the model improvement.Quantitative and qualitative analysis is carried out to verify the feasibility of the scheme.The weakly supervised recognition model and the improved CycleGAN model based on recognition network in this paper give a new solution to the problem of natural scene image understanding,which is mainly based on Pascal VOC data set.