The embodiment of the invention provides a fine-grained detection method based on a multi-class multi-attention mechanism, which comprises the following steps of: establishing feature mapping, importing extracted features into the feature mapping, and performing calculating to obtain a channel output value representing each channel in an input image; determining screening weights for the channel output values step by step, and obtaining feature vectors corresponding to the input images in the convolutional neural network in combination with the obtained screening weights; and performing randomselection from the feature vectors, constructing a sample set based on the attention category based on a selection result, constructing a minimum loss function corresponding to the sample set, and performing optimization processing on the minimum loss function to complete fine-grained detection. Attention perception features are extracted through an extrusion type multi-excitation module, the module is supervised by MAMC loss, and the loss enables a positive sample to be close to an anchor point and pushes away a negative sample at the same time. The method does not need a bounding box or partial annotation, and the processing steps are simplified.