Link prediction (LP) in social networks is to infer if a link is likely to be formed in the future. Social networks (SN) are ubiquitous and have different types such as human interaction and protein-protein networks. LP uses heuristic methods including common neighbors and resource allocation to find the formation of future links. These heuristics are sensitive to different types of social networks. Certain types of heuristics work better for some SN types, but not for others. Selecting the appropriate heuristic method for the SN type is often a trial and error process. Recent ground-breaking methods, WLMN and SEAL, demonstrated that this selection process can be automated for the different types of SN. While these methods are promising, in some types of SN they still suffer from low accuracy. The objective of this paper is to address this weakness by introducing a novel framework called PLACN that incorporates the analysis of common neighbors of nodes on target link and a combination of heuristic features through a deep learning method. PLACN is driven by a new method to efficiently extract the subgraphs for a target link based on the common neighbors. Another novelty is the method for labeling subgraphs based on the average hop and average weight. Furthermore, we introduce a method to evaluate the approximate number of nodes in the subgraph. Our model converts link prediction to an image classification problem and uses a convolutional neural network. We tested our model on seven real-world networks and compared against traditional LP methods as well as two recent state-of-the-art methods based on subgraphs. Our results outperformed those LP methods reaching above 96% of AUC in benchmark SNs.