| For Convolutional Nueral Network(CNN),this paper has presented a theoretical explanation and done two application researches.In theory,transforming the “interpretability of the deep learning” into the“friendliness of basis functions of the approximate function for human beings” is proposed,and a pattern of Transform Approximation is proposed,and the CNN is explained from this view.By adding artificially designed data node transforms,the traditional approximation methods can be turned from complicated to simple,from difficult to easy,and from inapplicable to applicable.As a variant of the traditional fully connected neural network,CNN essentially introduces some data node transforms,such as convolution,pooling,ROI Pooling and region proposals in some related frameworks.The first application research is proposing a SSD model whose loss parameter is set according to the characteristics of the class.The SSD model is a multi-class detection framework based on a convolutional neural network,whose loss function is formed by weighting and summing the classification error and the regression error.In the existing SSD model,the weights are same for different object categories.However,different categories may vary greatly.So this paper proposes setting different weights according to different object categories.A rank weighted method is proposed to set the weights.Finally,the comparative experiments are carried out on data sets,and results show that the performance of the improved model is a little better.The second application research is proposing an auxiliary driving related detection based on Faster R-CNN.Faster R-CNN is applied to the detection tasks related to auxiliary driving,including pedestrian detection,vehicle detection,traffic light detection and traffic sign detection.Collect related data sets,train and test.The results show that the Faster R-CNN can deal with the detection tasks well,and can detect target objects accurately,and can also achieve a good application effect in the actual scene.Figure 23;Table 4;Reference 90... |
