Research on Image Recognition and Deep Learning Models in Computer Vision

Abstract

The current mainstream deep learning models still have deficiencies in multi-scale target recognition and complex scene processing capabilities, which affects the generalization and robustness of the models. A multi-scale adaptive attention network (MSA-Net) is proposed for optimizing the performance of image recognition by combining multi-scale features with adaptive attention. MSA-Net extracts multi-scale features through a multi-branch convolutional network to capture target information of different resolutions, and introduces an adaptive attention module to dynamically optimize the global and local attention distribution, thereby enhancing the model's ability to focus on key areas. Experimental simulations are based on public datasets such as ImageNet and CIFAR-10. The results show that MSA-Net has significantly improved classification accuracy, with a classification accuracy of 94.6% on the CIFAR-10 dataset, which is 3.2% higher than ResNet; the mean average precision (mAP) is increased to 82.5%, and the inference time is optimized to 22.8 FPS. Visualization experiments verify the significant advantages of MSA-Net in complex backgrounds and multi-scale target scenes. Practical applications show that MSA-Net can achieve efficient image recognition in the fields of security monitoring, medical imaging, etc., and provide strong technical support for the automation and intelligence of related scenarios.