Orthopantomograms are essential tools for dentists in analyzing dental conditions like dental caries. Typically, dentists diagnose cavities by manually examining these images. However, manual detection can sometimes miss cavities due to factors such as negligence from excessive workloads and lack of expertise. Hence, the demand arises for computer-based smart vision technologies that employ machine learning, image processing approaches and artificial intelligence techniques to tackle these challenges. This study leverages an advanced deep CNN based DeepLabV3+ techniques such as atrous convolutions and encoder-decoder architectures to capture rich contextual information and fine-grained spatial details from orthopantomograms (OPGs). The encoder-decoder architecture enhances the ability to capture minute details in the segmentation of tooth caries, by capturing multi-scale context with multiple parallel filters employing different atrous rates, enabling the model to consider features at various scales simultaneously. The OPGs were obtained from the Nanded Rural Dental College & Research Centre in Maharashtra. A total of 153 OPGs selected and processed by an experienced dentist to identify the caries regions in the original images. The mask dataset is generated by extracting the caries region for the segmentation process. Various augmentation approaches such as zooming, rotation, horizontal and vertical flipping were applied to expand the dataset and achieved improved segmentation results. The results indicate that the model is highly effective in diagnosing dental caries, achieving an accuracy score of 99.2%, IoU score as 99.3% and Dice coefficient as 99.68%, respectively.