Abstract: In order to solve the problem that the measurement sensitivity of laser speckles decreases when the water temperature is higher than 20 ℃, a laser speckle image recognition and detection method based on depth learning is proposed. The speckle image data sets of 20.1 ℃, 20.2 ℃, and 20.3 ℃ were constructed. A multi-scale convolution neural network was used, combined with appropriate loss function and data enhancement technology, to optimize the characteristics of laser speckle images. Through the training and testing experiments of deep learning models on speckle datasets, high accuracy recognition of underwater temperature information speckle images was achieved, solving the problem of decreased sensitivity in contrast saturation measurement. The experimental results show that compared with AlexNet, VGG, and ResNet models, the accuracy of the GoogleNet model in underwater temperature recognition of speckle images reaches 99%. This study provides theoretical support for the in-depth understanding of temperature field distribution and its impact and provides valuable reference for related application fields.