IMAGE SALIENCY DETECTION METHOD BASED ON ADVERSARIAL NETWORK

摘要

Disclosed is an image saliency detection method for generating a convolutional neural network model by means of adversarial training, falling within the field of computer vision and image processing. The method comprises data pre-processing, network construction, selecting a suitable parameter, and using a stochastic gradient descent method and an impulse unit to perform training. The data pre-processing is pre-processing the collected mass data and tags. The network construction is designing a network structure and a specific kernel function. Selecting a suitable parameter comprises a learning rate, a momentum factor and the number of images crammed into a network each time. Using a stochastic gradient descent method and an impulse unit to perform training reduces the possibility of network overfitting. By means of the method, a saliency map can be obtained more accurately.