In this paper, three convolutional neural network models are used to achieve end-to-end recognition of blood cell images. The network model parameters are initialized by transfer learning from the pre-trained model on ImageNet, and then the blood cell images are input into the model, and the network model training is completed by back-propagation to continuously update the parameters. For small-scale datasets, the number of blood cell images is expanded using data increments to improve the generalization ability of the model. Experimental results on the BCCD dataset show that the best result MobileNetV2 achieves an accuracy and precision of 0.894 and 0.916, respectively.

To address the current problem of frequent piracy infringement of digital images, this paper proposes an image watermarking algorithm based on DWT and DCT. Firstly, the binary watermark is encrypted by the Logistic Chaos algorithm, which produces a chaotic sequence with better uncertainty and initial value sensitivity; secondly, the Arnold transform is applied to the encrypted watermarked image to eliminate the correlation between the watermarked pixels, so that the robustness of the image information is improved; the original image is decomposed by the second-level wavelet in the embedding process, and then the low-frequency sub-band LL1 is blocked In the embedding process, the original image is decomposed by the second-level wavelet, and then the low-frequency sub-band LL1 is chunked, and the pre-processed watermark information is embedded by modifying the IF coefficients through DCT transform. In this paper, Matlab simulation experiments are conducted for this algorithm, and the experimental results have good invisibility and good robustness when subjected to various malicious attacks, which is practical for safeguarding digital image copyright.