Image Classification Method Based on Linear Superposition Features and Convolutional Neural Networks

To solve the engineering skills problem that traditional Convolutional Neural Networks(CNNs) is of long training time,an image classification algorithm which makes linear superposition features as the input of CNNs is proposed,on the basis of CNNs has strong spatial information characteristics. In this algorithm,mainly researches the relation of derivative with respect loss function to weights between the output of CNNs with original image and LS features as its input,and analyses update mechanism of connection weights. The classification results of experiments on the MNIST database of handwritten digit database show that CNNs with LS features as its input not only reduces the training time of the model markedly but also requires no complex engineering tricks. It is thus concluded that the proposed method is viable in the image classification.