In this study, two novel U Net–based architectures were proposed and evaluated to improve image segmentation performance on the Oxford IIIT Pet dataset. In the first proposed architecture, two decoder paths and three bottleneck blocks were designed to enhance the network’s capacity for learning complex features. In the second proposed architecture, two parallel decoder paths with direct skip connections from the encoder were employed to achieve better extraction of low level features. For performance evaluation, architectures were trained from scratch. The accuracies of the U Net, the first proposed architecture, and the second proposed architecture on the training set were 0.9154, 0.8654, and 0.9270, with corresponding loss values of 0.2112, 0.3433, and 0.1825, respectively. On the test set, these architectures achieved accuracies of 0.8763, 0.8611, and 0.8906, along with loss values of 0.3881, 0.3654, and 0.3414, respectively. These findings indicate that the second proposed architecture outperforms both the standard U Net and the first proposed architecture. The second model demonstrates higher accuracy and lower loss, attributed to the improved decoder pathways and redesigned skip connections.