To effectively solve the problem of poor detection effects of remote sensing image target detection algorithm in complex background,an improved YOLOv4 target detection algorithm was proposed.A cross-stage residual structure was designed to replace the simple residual structure in the CSP module of the original backbone network to reduce the amount of model parame-ters and computational burden.The CBAM attention mechanism was introduced to strengthen the interaction of effective features between CSP modules.The cross-stage hierarchical convolution module was used to reconstruct the processing method of the deep feature map in the feature fusion stage to prevent network degradation and gradient disappearance.The Mish activation function was used to enhance the extraction ability of the fusion network for nonlinear features.Experimental results on the RSOD and DIOR data sets show that the test mAP of the improved YOLOv4 algorithm is 4.5％and 7.3％ higher than that of the original YOLOv4 algorithm,and its detection speed reaches 48 fps and 45 fps respectively,which ensure the real-time per-formance and greatly improve the detection accuracy.
维普
万方数据