Fully Connected to Fully Convolutional: Road to Yesterday

In the last decade, several convolutional networks have been developed for the semantic segmentation, which have shown excellent performance in recognizing and labeling objects in images. Most of these networks involve large-scale architectures that can detect tens or hundreds of predefined classes. With the exception of fully convolutional networks, most applications use architectures that, after convolutional layers, use a common classifier to classify the extracted features. In this paper, the method of converting a network, which as classifier, has two flatten and dense layers (fully connected), to a fully convolutional network is described. The main advantage of this method is the ability to work on inputs of variable size and produce an output map instead of a number, which is the advantage of fully convolutional networks. Newer models of the Deep Learning area generally use training images in which areas of interest are determined by masks; but in the proposed method only labeled images are given to the network. The details of the proposed method are expressed in the form of a new problem of classification of boards with calligraphy of Shekasteh-Nastaliq and Suls, and classification of apple leaf diseases (as two-class problems) and the problem of identifying hand written Persian digits. For this purpose, first a convolutional network with the last fully connected layer is designed and trained for square images. Then a new fully convolutional model is defined based on the previous model and the weights of the previous model are fed to the new model. The only difference between the two models is in the last layer, but the new model will be able to work on input images of any size. Experimental results show the efficiency of the proposed approach.