The invention discloses a navigation
underwater target detection method based on multi-
modal fusion and adversarial recognition, and relates to the field of
computer vision, and the method specifically comprises the steps: obtaining visible light,
sonar and hyperspectral data of a navigation platform, completing the cross-
modal space-
time alignment, and generating a synthetic sample through text driving to supplement training data; carrying out
inversion recovery on the low-light image by adopting a
light transmission model; constructing an end member matrix to complete hyperspectral unmixing, and obtaining an abundance matrix through physical constraint enhancement; packaging the multi-
modal data as a training tuple, generating a disturbance sample, and optimizing the model through a joint
loss function; in the reasoning stage, a real-time environment is adapted by prompting vector self-calibration and multi-view uncertainty weighting fusion, abnormal detection and
mode switching are realized in combination with a prediction entropy value, and finally a result is packaged and data is returned to form an iterative optimization
closed loop. According to the method, the target identification degree is improved through multi-modal feature complementary fusion, the anti-interference capability is enhanced by means of adversarial training, and a complex
underwater environment is adapted based on dynamic reasoning and closed-
loop optimization.