Research on CNN for Anti-missile Object Detection Algorithm Based on Improved Attention Mechanism

In military applications, it has great research value to mine reliable military information from remote sensing image data. Since the precision and recall rate of the CNN (convolutional neural networks) in the object detection process is low, and cannot meet the requirements of military combat indicators. An improved CNN for the anti-missile object detection algorithm based on improved attention mechanism is studied. First, DehazeNet algorithm was used to improve the recognition of the target while is obscured by the fog in the antimissile remote sensing data set, and the data set was re-clustered through the K-means algorithm; then based on the idea of ECANet attention mechanism to improve and reorganize its pooling layer, and an improved attention module I-ECANet (Impproved ECANet). Improved the last residual layer of each group of convolution and residual layers in the YOLOv3 network to the I-ECANet attention module, training network before and after improvement; finally, use the test set to test and analyze the anti-missile remote sensing image. The results shown that the precision, recall, and average detection precision of the algorithm have been greatly improved, with the precision rate reaching 97.6%, an increase of 6.8%, and recall rate reached 95.7%, an increase of 3.8%, Average detection precision reached 95.1%, an increase of 4.4%, and detection speed reached 33.87 images per second. It shows that the improved algorithm proposed in this paper has wide application value.