The invention discloses a convolutional neural network loss measurement method based on regional characteristics. The method comprises the following steps: 1) generating a corresponding regional characteristic graph for each category in the last layer of a convolutional neural network; 2) reducing the regional feature map generated in the step 1) through an average pooling method; 3) carrying outcombined training on the local classifier and the global classifier by adopting a convolutional neural network loss measurement method based on the regional feature map in the training stage; and 4) combining the local classifier into the global classifier by utilizing weight parameters obtained by training in the testing stage to obtain a final classification result. According to the method, theexpression capability of the network to a local area is obviously enhanced. Parameters in the neural network are reduced, and the calculation complexity of the network is reduced, so that the generalization capability of the network is enhanced, and local loss and global loss are effectively fused, so that the performance of the convolutional neural network is obviously improved.