Research On Image Generation And Velocity Estimation Based On Generative Adversarial Networks

In recent years,flow velocity measurement has played an important role in the regulation of hydropower resources.Our country has developed a large number of rural small hydropower stations for real-time hydropower dispatching.The actual natural environment is complex and variable,the river flow rate in that tends to change with the seasons,time and terrain.Although the current methods have achieved good results,there are still deficiencies.On the one hand,the traditional manual flow velocity measurement method consumes a lot of manpower and material resources in actual operation,and the safety of professional surveyors cannot be reliably guaranteed in the event of flooding or complex terrain.On the other hand,the popular particle image velocimetry method probably needs to throw suitable tracer particles into the river,and the selection of the tracer particles has such conditions as environmental protection and feasibility.Therefore,the design and development of a new non-contact flow velocity measurement method is of great significance for the remote areas on the premise of saving labor costs and equipment costs as much as possible.With the advent of big data and deep learning eras,the classification and prediction of big data magnitude has begun to be applied to different actual scenarios.Thanks to its powerful self-learning ability,deep learning has played a pivotal role in the field of pattern recognition.Starting with the advantages and disadvantages of the existing popular flow velocity measurements,this paper designs and explores a new non-contact flow velocity estimation method by using the generative adversarial network and convolutional neural network which are excellent in image generation and classification.The main work is as follows:(1)Aiming at the problem of insufficient richness of river surface texture features as well as realizing the data enhancement effect,a conditional generative adversarial network algorithm based on boundary balance is proposed to realize water flow image generation in normal sunny weather.The label information is introduced to forcibly guide the generated data direction of the generated network so as to reduce the degree of freedom in data generation.To ensure the reliability of the generated sample,in addition to the authenticity judgment of the discriminator,a verification module is added to achieve higher quality water flow image generation.The final experiment verifies the effectiveness of the proposed algorithm.(2)On the basis of the above research,in order to increase the abnormality of data set and realize the simulation of water flow image with fog obscuration,a generative adversarial network based on improved Pix2 Pix is proposed.A new target loss function is introduced without reducing the generation sample ambiguity,alleviating the over-fitting problem on the water flow data set,thus enhancing the robustness of the subsequent classification network.Finally,the water flow data is simulated and generated by experiments.(3)In order to realize the classification of flow velocity at different flow rates,in addition to enhancing the dataset,a classification model is proposed in this paper.A convolutional neural network based on multi-feature fusion is proposed to realize the features fusion and learning.The algorithm is to enhance ability of the classification network which distinguishes the different river images at different flow rates with small differences.The experiment results verify the effectiveness of the classification network for different flow velocity and different resolutions on the server platform.In summary,this paper mainly studies the image generation of river surface and flow velocity estimation based on generative adversarial networks.In addition,this paper applies the proposed algorithm to the actual small hydropower dataset,and achieves relatively good experimental results.