High Precision Target Detection Algorithm Based on Dual Channel Attention Perception Fusion

Abstract

With the development of modern technology and the application of the times, the accuracy of high-resolution image target detection is gradually improved. In this paper, a multi feature extraction and multi feature fusion network is proposed for the loss function of feature extraction, feature fusion and uneven positive and negative samples. In order to solve the problem of multiple features in multi feature extraction, combined with the two-dimensional CA spatial attention mechanism and residual network, the global features are preserved and sensitive information extraction is strengthened; Aiming at the problem of feature redundancy in multi feature fusion, the improved bifpn structure is used to effectively solve the problem of feature redundancy after feature extraction; Aiming at the training model degradation caused by the imbalance of positive and negative samples, the focal loss function is improved to effectively solve the model degradation caused by samples. The improved multi feature extraction and multi feature fusion network effectively improves the accuracy of target detection. The improved algorithm has the highest accuracy in dota data set, and the test accuracy is 78.98%.