The semantic segmentation process divides an image into its constituent objects and background by assigning a corresponding class label to each pixel in the image. Semantic segmentation is an important area in computer vision with wide practical applications. The contemporary semantic segmentation approaches are primarily based on two types of deep neural networks architectures i.e., symmetric and asymmetric networks. Both types of networks consist of several layers of neurons which are arranged in two sections called encoder and decoder. The encoder section receives the input image and the decoder section outputs the segmented image. However, both sections in symmetric networks have the same number of layers and the number of neurons in an encoder layer is the same as that of the corresponding layer in the decoder section but asymmetric networks do not strictly follow such one-one correspondence between encoder and decoder layers. At the moment, SegNet and ESNet are the two leading state-of-the-art symmetric encoder-decoder deep neural network architectures. However, both architectures require extensive training for good generalization and need several hundred epochs for convergence. This paper aims to improve the convergence and enhance network generalization by introducing two novelties into the network training process. The first novelty is a weight initialization method and the second contribution is an adaptive mechanism for dynamic layer learning rate adjustment in training loop. The proposed initialization technique uses transfer learning to initialize the encoder section of the network, but for initialization of decoder section, the weights of the encoder section layers are copied to the corresponding layers of the decoder section. The second contribution of the paper is an adaptive layer learning rate method, wherein the learning rates of the encoder layers are updated based on a metric representing the difference between the probability distributions of the input images and encoder weights. Likewise, the learning rates of the decoder layers are updated based on the difference between the probability distributions of the output labels and decoder weights. Intensive empirical validation of the proposed approach shows significant improvement in terms of faster convergence and generalization.
