Image Recognition using Improved ResNet50 Model with Novel Attention Mechanism

Abstract

The process of classifying an image is diverse and subject to various factors and it typically require a deep neural network. In this paper we presented a new approach to improve the existing deep learning model by including an attention mechanism. We proposed integrating an attention mechanism (CBAM) convolutional block attention module with the baseline ResNet50 architecture to enhance the accuracy of the model. We trained the Intel Image Classification dataset with and without the Attention Mechanism, which has about 25k images of resolution 150x150 divided among 6 categories. (CBAM) ResNet50 model achieved accuracy of 0.88, which is considerably better than the accuracy of the existing ResNet50 model (0.84) tested on the same dataset. Deep transfer learning has earned a lot of attraction in the field of image recognition, and these findings show that optimizing models for specific problems might be important and improving prediction accuracy.