Aiming at the problem of low accuracy and poor real-time performance of commonly used target detection algorithms, a novel ship target detection algorithm based on convolutional neural network with feature fusions is proposed to detect multi-scale ship targets in complex scenes. The proposed method inherits the network structure of SSD and introduces the deconvolution feature fusion module and the pooling feature fusion module into it to generate the new feature maps with richer semantic information for both ship classification and boxes regression. In addition, we used a focal classification loss function in the training strategy to deal with the imbalanced difficult and easy samples in the training process. The experiments tested on the ship detection dataset demonstrate that the proposed method shows a better adaptability to ship detection of different sizes in complex scenes. On the extended experiment, the proposed method performance over SSD in blurry object detection.