To address the issue that instance segmentation models often prioritize speed over accuracy, leading to frequent detection errors or mask omissions, we propose a novel Image Segmentation founded on hybrid dilated CNN and optimized IoU. Firstly, we enhance the backbone network using a hybrid dilated convolutional neural network to improve feature extraction capabilities, mitigating the loss of continuity information. Secondly, we introduce an objection localization and segmentation loss function built upon an optimized IoU metric. This algorithm ensures that the optimized IoU does not set to zero even when objects do not overlap, thereby addressing the challenge of gradient optimization. Additionally, we employ a Soft-NMS algorithm to enhance object localization accuracy and resolve issues related to mask omissions. Finally, we utilized the MS COCO dataset and conducted experiments under an Intel Core i7-9900K CPU and NVIDIA Titan RTX GPU environment. Our proposed algorithm has proven to be superior to prevalent instance segmentation models like FCIS, Mask R-CNN, and YOLACT, showing better performance in both speed and accuracy, as indicated by the results. Specifically, comparing to the FCIS model, FPS increased by 6.4, and the accuracy rate improved by 6.3%; compared to the Mask R-CNN model, FPS enhanced by 27.1, and the accuracy rate reduced by 0.4%; compared to the YOLACT model, FPS decreased by 2.3. The average detection and segmentation time for our algorithm was 29.9 ms, with an instance segmentation accuracy of 96.7%. Overall, our proposed algorithm offers superior performance, effectively balancing the need for real-time processing with high accuracy in instance segmentation tasks.