An Object Detection Algorithm of Literatures’ Key Information Based on Optimized Cascade R-CNN

To address the problems of relying on electronic repositories and being vulnerable to network influence in obtaining key information of literature in mainstream literature management software, an optimized Cascade R-CNN-based algorithm is proposed for detecting key information of literature. Based on the cascade regional convolutional neural network (Cascade R-CNN), a Res2Net feature extraction network embedded with SimAM and CBAM attention modules was used to acquire features at different scales. The bi-directional feature pyra-mid network (BiFPN) is introduced. The BiFPN enhances the feature represen-tation capability by adding residual links and weight fusion to better balance the feature information at different scales; improves the algorithm’s ability to distinguish hard samples by using the focal loss function to balance the unbalanced number of positive and negative samples; and optimizes the anchor mechanism to accelerate the network convergence. The experimental results show that the optimized Cascade R-CNN object detection algorithm has the best comprehensive performance, and the average accuracy (AP<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">0.5</sub>) reaches 95.5%, which is 3.5% better than the original model, and takes 7.95 FPS to achieve better detection results.