At present,facial expression recognition methods have the problems of large number of parameters,large consumption of computing resources and low recognition accuracy.Aiming at the above problems,a lightweight human facial expression recognition method based on conditional coordinated attention mechanism is studied.First,the number of layers of MobileNet V3 network is reduced,while the numbers of intermediate channels and output channels of the inverse residual structure are increased to 1.5～3.2 times of the original number.Mish is used instead of Hardswish activation function to realize the nonlinearization after feature extraction.Secondly,an improved coordinated attention mechanism is introduced to encode the tensor information embedding along horizontal and vertical directions sequentially by maximum pooling and average pooling.And tensor information integration is used to generate features with global sensory field and precise location information to extract detailed information of facial expressions in space and channel location.Finally,experiments are conducted on the publicly available datasets FERPlus and RAF-DB,and the results show that the proposed method reduces the number of parameters by 15.91%,and the accuracy rates are 88.84%and 85.90%,respectively,which are 0.83%and 1.39%higher than the accuracy rates of the model before improvement.The method has good recognition performance and validate the effectiveness of the proposed method.
国家科技期刊平台
NSTL
万方数据
