At present, the detection of surface defects on the market often relies on human visual recognition. This traditional method consumes a lot of human and material resources, which hinders the development of the market to a certain extent. At the same time, in recent years, deep learning technology has been used in image recognition. In the development of classification, more and more fields are using deep learning for object detection. In order to solve the problem of high miss rate of surface crack defects in traditional manual detection methods, a method based on deep convolution neural network DenseNet (Densely Networks) to detect crack surface defect data was proposed, and two tests were performed on the network The improvement is that one is to add a coding-decoding structure so that the network input and output are graphs, and the second is to increase the number of network convolution kernels and increase the receptive field size of the network feature extraction, so that the improved DeepDenseNet (Deep Densely Networks) The network converges faster, the ability to learn features is stronger, the detection effect is better, and the detection results are compared with the deep convolutional neural network VGGNet model. Experimental results show that the proposed method is effective for small sample surface crack defect data Set has a good detection effect.