Research On Filter Visualization Of Visual Geometry Group Model

In recent years,in-depth learning has achieved great success in machine vision,like object identification and detection,natural language processing and other fields,which can be regarded as a big stride development in promoting artificial intelligence and solved many complex pattern identification tasks.However,in-depth learning models are often referred to as "Black Box",namely the representations learned by the model are difficult to be extracted and presented in a human understandable way,which greatly limits the development of in-depth learning,especially in fields like automatic driving,finance and medicine where in-depth learning are used for key decisions making for great reliability of algorithms is often needed.In visual processing,the most valuable Convolutional Neural Network can allow us to understand the in-depth learning algorithm through visualization,and to further explore the "Black Box".Since 2013,scientific researchers have proposed many techniques or methods to visually explain the "Black Box" of Convolutional Neural Network.Visualizing the filter of the Convolutional Neural Network is one of the methods.However,due to the different response values of different filters to different images and the effect of Relu activation function on the “inactivation” of neurons,the gradient of the in-depth learning model using Relu activation function is zero when it propagates back,resulting in the failure of this method to fully visualize the filters in the Convolutional Neural Network.This paper presents two effective methods to enable complete and effective visualization of every filter.In this paper,the filters of the trained VGG16 model are visualized to demonstrate the method as follows:Method I: Replace the single overall Relu activation function training with Leaky Relu activation function training for several times and then change the activation function to Relu activation function training.Method II: The randomly generated images that cannot respond to the high-level filter will first respond to the low-level filter through the recursive method,so that the randomly generated images will move in the direction of the response of the filter to be visualized,and the visualization task of the filter will be completed.One of the methods to solve the neuron inactivation caused by the Relu activation function is to improve the activation function,such as replacing the Relu activation function with the Leaky Relu activation function.On this basis,this paper compares the three filter visualization methods(the activation function as Leaky Relu,method I and method II),and then analyzes the advantages and disadvantages of the two methods proposed in this papers;Finally,to lower the threshold of VGG16 model filter visualization,a graphical user interface(GUI)software,VGG16 visualization algorithm,is developed and introduced.