In recent years, object detection technology has developed rapidly, especially the FPN has been widely adopted in production application and academic research. Compared with the traditional network, the accuracy of EfficientDet based on the FPN structure has reached 55.1% on the COCO dataset, and the number of parameters and computation cost is also greatly reduced by 9–25 times. However, the EfficientDet training process does not integrate the upper and lower level feature information quickly enough, especially depending on the multi-layer BiFPN overlap and increasing input size and channel number, therefore, we proposed WavesNet based on EfficientDet. First, we integrated seven channels on the basis of BiFPN to make all levels of the fusion more quickly and effectively, and then applied the ComC (Combine Convolution) extraction method in the initial phase of feature extraction. In addition, the EfficientDet fast normalization formula was improved to ensure the numerical stability of the features. Experimental results show the WavesNet has 30% faster convergence rate and 0.5% higher accuracy on COCO dataset than the EfficientDet. The detection effect on small targets is obvious.