Haze scenes seriously affect the working performance and accuracy of computer vision systems. As an important research direction in the field of computer vision, image dehazing has always attracted the attention of researchers. Convolutional neural networks play a good role in image processing problems by virtue of their advantages. Therefore, convolutional neural networks are also used in image dehazing tasks. The mainstream dehazing algorithms are mainly divided into two categories, one is the image recovery algorithm based on atmospheric scattering model, and the other is the training learning dehazing algorithm based on convolutional neural network. Although the recovery class of dehazing algorithms considers the nature of haze image formation and obtains good results, the pathological nature of the atmospheric scattering model leads to the need for precise a prior conditions and harsh constraint rules, making the applicability of this class of algorithms limited. The idea of convolutional neural network-like dehazing algorithm is to train a convolutional network model with dehazing capability on synthetic dataset. In recent years, some researchers have designed a variety of image dehazing networks, although all of these networks achieve the effect of image dehazing, they still have many shortcomings. The main manifestation is that the dehaze image is too dark, the detail is lost seriously, the color is distorted and the dehazing is not complete. To address these problems, an image dehazing algorithm based on step-type network extraction and attention cross-fusion mechanism is proposed. The whole network model contains three modules, the stepped feature extraction network, the feature fusion module based on the attention mechanism and the clear image generation module. Among them, the step-type network performs detail and contour feature extraction of haze images, the fusion module adaptively fuses the detail and contour features in an attention mechanism, and the generation module outputs the dehaze images. In the feature fusion module, the residual structure is introduced to enhance the feature information and improve the accuracy of the network. The loss function used for network training is a combination of mean square error loss and perceptual loss, and the perceptual loss can effectively improve the semantic information of the features with haze images, which in turn leads to a more accurate dehaze image. The network model is considered to reach stability after the loss values reach convergence. After the network model is trained, rich experiments are used to demonstrate the validity and feasibility of the proposed model. The experiments in this article include two parts: the main experiment and the ablation experiment, and both the main experiment and the ablation experiment are analyzed in comparison from two perspectives: subjective evaluation and objective evaluation. The subjective evaluation uses experimental objects with haze images in real environments and synthetic images in datasets, and the objective evaluation uses some publicly available and widely used quantitative metrics. The experimental results show that the proposed model has good results for both haze images in real environment and synthetic images in the dataset. The dehaze image obtained by the proposed model has richer detail information, more natural color effect, more suitable brightness information and more complete dehazing effect. Experiments on different datasets demonstrate the wide applicability of the proposed model. In the objective evaluation, the proposed model also shows a clear advantage. It has a clear lead in the no-reference metric visible edge increase rate, average gradient, number of saturated pixel points and histogram similarity, and also outperforms the comparison algorithm in structural similarity and peak signal-to-noise ratio. The main experiments demonstrate the validity and feasibility of the proposed model, and in addition, local detail comparison experiments are used to demonstrate the performance of the proposed Moses on the detail information of the dehaze images. To demonstrate the necessity and importance of each component module in the proposed model, ablation experiment is used in this paper. The ablation experiments demonstrate the effectiveness of the step-type network for extracting detail features and contour features, and the effectiveness of the fusion approach under the attention mechanism. Although the proposed model obtains better dehazing effect, it is weaker for dense haze images. The dehazing method for dense haze images is something that needs to be focused on in the future.