Research on Optimization of Object Detection Technology Based on Convolutional Neural Network

Abstract

Aiming at the optimization problem of object detection technology, an instance segmentation model algorithm based on the improved Cascade Mask R-CNN is proposed. Firstly, information fusion is performed on the mask branches of each cascade stage, so that the mask information of the current stage is jointly determined by the segmentation results of the previous stage and the current stage. Secondly, GA-RPN is used to replace the original RPN to generate high-confidence and low-density proposal regions. At the beginning of the cascade, the IoU threshold is increased for training, thereby improving the performance of the network model. Finally, the model of this paper is implemented, and compared with the official implementation on the public data set, which verifies the efficiency of improved algorithm.