Distribution of fractures, stratigraphy and grains of conglomerate can be analyzed visually in complex reservoirs such as carbonate and sand conglomerate by electrical imaging logging. A fully convolutional deep neural network model is proposed in this paper which captures large volumes of bottom prior statistic features in image and realizes gaps filling in electrical logging images in wells with large caliper to form full borehole covering images by gradually optimizing parameters of neural network model without large number of learning samples. Compared with traditional encoder-decoder model, skip pattern is utilized to connect output of encoder layers with corresponding decoder layer which is helpful for recover local details in images and atrous convolution is adopted to capture multi-scale contextual information. Experiments show that mean error of gray level of pixels is decreased by about 12% and gaps filling effect for images with complex lithology is better in this paper compared with mainstream gaps filling algorithms.