Design Of Optimization Scheme For Deep Learning Model In Image Field Under Visual Guidance

Deep learning has developed rapidly in recent years,has been widely used,and has achieved many remarkable results.Under this background,there are more and more researchers engaged in deep learning research,but most researchers face the problem of difficulty in tuning models.The parameter tuning of the model is mainly divided into two parts:the adjustment of the network structure and the adjustment of the training hyperparameters.Each adjustment has many possibilities.Usually,developers tune in based on previous tuning experience and make a lot of tentative adjustments.When the data scale is large and the network model is deep,each training consumes a lot of time.Therefore,it is very important to find an effective method to optimize the model.Convolutional neural networks are an important branch of deep learning,and they also face difficulties in tuning.In response to this problem,this paper proposes a visual aided tuning method for convolutional neural network models.The main objects of visualization are the output and parameters of the middle layer,which implements dif-ferent visualization methods in two-dimensional or four-dimensional data dimensions.For two-dimensional data,use heat map rendering to visualize;for four-dimensional data,slice along certain two dimensions to form an image matrix for visualization.The specific implementation method depends on whether the data is output or parameters.This paper uses the cosine distance to measure the similarity of different data in a cer-tain dimension,combined with the hierarchical clustering method,and draws the results using the hierarchical clustering tree,which helps people to analyze the model more in-tuitively.In addition,by performing kernel density estimation analysis on the entire data or along a certain dimension,and drawing its probability density distribution,the effect of the model can be indirectly analyzed.By visualizing the results,locate the network layer or neuron that caused the model failure,and give appropriate tuning sug-gestions.The paper verifies through experiments that the proposed method can reduce the blindness of researchers in model tuning and improve development efficiency.