Research on High-Precision Object Detection and Instance Segmentation Using Mask-RCNN

As the scope of target detection applications continues to expand, the application of the Mask-RCNN model in object detection and instance segmentation tasks has become a focal point of research. This paper provides an in-depth analysis and optimization of the Mask-RCNN model, implementing a series of improvements that significantly enhance its performance. Firstly, we scaled the images to 1024x1024 pixels, thereby increasing the model’s capability to process high-resolution images. Secondly, by employing a dynamic bounding box generation method, the model’s target localization accuracy is greatly improved. Furthermore, the utilization of a smaller learning rate and gradient clipping techniques effectively enhances the model’s convergence speed and stability. The optimized model demonstrated outstanding performance on the MS COCO 2017 dataset, achieving a mean Average Precision (mAP) of 0.833. This result highlights the model’s superior performance in object detection tasks, particularly in terms of high precision and recall rates. Additionally, the improved model exhibits high stability and consistency in classification accuracy and bounding box regression prediction capability. Experimental results indicate that the Mask-RCNN model not only holds significant theoretical research value but also demonstrates substantial practical utility, providing robust technical support for object detection and instance segmentation tasks in real-world applications.