Convolutional neural network target classification method based on novel loss function

The embodiment of the invention provides a convolutional neural network target classification method based on a novel loss function. The method comprises the following steps of: introducing a dynamicangle-adding margin which is changed along with a target class characteristic center vector and each subclass characteristic center vector in a model training process; calculating a first cosine valuevector of an included angle between a feature vector of a sample image and a feature center vector of a target class; according to the dynamic angle adding allowance and the included angle between the target class characteristic center vectors; obtaining a second included angle through a second preset formula, cosine of the second included angle is calculated; obtaining a second cosine value, replacing the cosine value of the included angle between the feature vector in the first cosine value vector and the feature center vector of the target class with the second cosine value; and calculating the loss of the model ac