Maritime object detection and segmentation are vital for navigation, surveillance, and environmental monitoring, particularly with the increasing use of Unmanned Surface Vehicles (USVs). The maritime environment poses challenges such as dynamic backgrounds, complex textures, and adverse conditions. In the proposed study, we propose an improved semantic segmentation approach using DeepLabV3+ with ResNet-101 and z-score normalization. The model was trained on a dataset of 7815 images, organized into image and semantic mask folders. Training spans 80 epochs, with the performance evaluated by pixel accuracy and Mean Intersection over Union (IoU). The results show the effectiveness of the model, achieving a validation loss of 0.767156 and a validation accuracy of 97.89 %. Visualizations of the original image, predicted mask, and overlaid results demonstrated accurate object segmentation in complex maritime environments. The proposed approach outperformed standard DeepLabV3 models and other state-of-the-art methods in capturing fine object details under challenging conditions, potentially enhancing various maritime applications through more effective and reliable autonomous systems.