The invention discloses an improved gesture image feature extraction method based on a DenseNet network. Acquiring a gesture to obtain an original gesture image; performing convolution downsampling through a convolution downsampling network structure, selecting feature tensors of a shallow layer and a deep layer, and inputting the feature tensors into a DenseNet-B module of a fused Drop-Path module to obtain two feature tensors; after fusion, obtaining a feature tensor of multi-scale feature fusion, compressing the feature tensor through a transition layer, and inputting the compressed feature tensor into a DenseNet-B module of the fusion Drop-Path module to obtain a feature tensor containing multiple scales and high dimensions; and obtaining a classification result through a global average pooling layer, a full connection layer and a softmax classifier. According to the method, feature tensors of different depths in a down-sampling network structure are included, large target objects and small target objects can be accurately recognized, meanwhile, a Drop-Path module is fused in the DenseNet network, the parameter quantity is effectively reduced while the precision is not reduced, the model training speed is increased, overfitting is prevented, and the gesture recognition accuracy is improved.