Aiming at the research shortcoming of Human Vision System (HVS) for Im age Quality Assessment(IQA),a method of IQA combing front perception and back internal generating of HVS characteristic is proposed.Gradient feature informat ion of every visual channel is deeply decomposed based on the back Internal Gene rative Mechanism (IGM) of HVS after gradient feature salience maps of all visual channels for an image are extracted based on front perception characteristic of HVS.In order to overcome pooling problem of multi-channel assessment,a model of visual multi-channel pooling for IQA based on BP neural network is construct e d and used to pool multi-channel gradient feature assessment of three visual la y er,respectively.Finally,model pooling assessment of every visual gradient fea ture layer is deeply pooled from the inner layer to the outer layer by designing adaptive regression algorithm.Experiment results show that the proposed poolin g model obviously improves the assessment indexes′ level of each objective algo rithm.Among them,the root mean square error (RMSE) is less than 3.4153,the Pe arson linear correlation coefficient (PLCC) is greater than 0.9900,and the Spea rman rank order correlation coefficient (SROCC) is greater than 0.9725.The asse ssment indexes of the proposed method hold higher level and better stability tha n the existing methods.