Abstract: Aiming at the new sliding puzzle captcha program with pseudo notch effectively preventing the attack of existing methods, this paper proposes an improved ResNet18 sliding puzzle captcha cracking method. In order to ensure the generalization of the training model, millions of training samples were obtained through data enhancement and images were preprocessed. The preprocessed images were sent to the improved ResNet18 for training and testing to obtain the network model. The model was used for slider detection and notch detection to calculate the distance between the slider and the notch, and the random curve fitting algorithm was used to generate the sliding trajectory. Selenium drag slider was used to complete the puzzle verification. The experimental results show that compared with the traditional ResNet18, the improved ResNet18 reduces the number of parameters by 41% and the GFLOPs by 59%. Under the condition that the detection accuracy is increased by 1.8 percentage points, the inference speed is 2.75 times faster, and the new and common sliding puzzle verification codes can be effectively solved. The mAP reaches 98.66%, and mAS reaches 3.68 s, with good universality and overall performance better than existing methods.