Neural network model input channel integration method for transfer learning

Abstract

The invention discloses a neural network model input channel integration method for transfer learning, and the method comprises: firstly modifying a data input layer description in a network model description file, deleting an original data input layer description, adding a data input layer description based on an image data list, and renaming a to-be-modified first convolution layer; reading a neural network model weight file, and modifying the neural network model weight file, so that the corresponding neural network model only needs single-channel data as input; and finally, carrying out real-time target detection, identification and test on the single-channel image data. According to the invention, network model weight files trained by three-channel training sample data are analyzed and integrated; a network model weight file input by a single channel is formed and directly deployed on an embedded platform, channel copying operation is not needed, redundant memory space does not need to be occupied, the calculation amount of a neural network is reduced, the pressure on information processing resources is reduced, and real-time target detection and recognition can be directly completed on single-channel data.

Description

technical field [0001] The invention relates to a neural network model input channel integration method for transfer learning, which belongs to the field of neural network intelligent target detection and recognition. Background technique [0002] The target detection and recognition algorithm based on convolutional neural network needs a large amount of training sample data in the process of model weight training. For some specific application environments, the amount of image data containing valid targets is too small to support algorithm model training, and it is necessary to use public databases for pre-training and other operations. However, most of the public databases are naturally collected three-channel color images, so the input of the trained model is also required to be three-channel data. However, most of the currently used visible light, infrared or radar SAR scene data are single-channel, which cannot meet the input requirements of neural network models. [...

AI Technical Summary

Problems solved by technology

However, the currently commonly used visible light, infrared or radar SAR scene data are mostly single-channel, which cannot meet the input requirements of neural network models.

[0003] In practical application, a simple and feasible method is to copy two channels of single-channel image data and splicing in the channel dimension to generate three-channel data. However, after testing, this method has inevitable defects: in the embedded information processing platform Copying the image channel on the network not only takes up extra memory space, but also increases the processing time of the copy channel, and also increases the amount of calculation of the neural network, which puts pressure on the already relatively tight information processing resources and reduces the overall target detection performance. Identify the real-time nature of system operation

Images