Date of Publication :29th March 2024

Abstract: Image Classification accuracy results are strongly based upon feature extraction methods adopted. Feature extraction is the dimensionality reduction process that efficiently represents only the meaningful parts of the im- age as a comparative lower dimensional feature vector. Traditional methods could not produce optimum results in case of remote sensing images due to much more complexity of remote sensing images as compare to normal images. Classification becomes tedious for remote sensing images as level of abstraction converges from pixel to objects. Thus traditional methods encoding color, texture and shape features proved to be inefficient for classify- ing complex remote sensing images. Scene categorization requires high level of features that are comparatively more representational than local and global features based on color, shape and texture. These traditional feature extraction approaches were drifted towards convolutional neural networks as these networks efficiently extracted abstract features. Pre-trained models employing transfer learning could produce superior results as compared to earlier traditional machine learning as well as neural networks based methods used for feature extraction and classification. Pre-trained models utilizes transfer learning in which learned features in form of weights for one general task is used for extracting features of a particularly specific task. Thus, transfer learning speeds up the training phase. This letter is to enrich the accuracy of image classification employing transfer learning with study of deep learning models models VGG-19, Inception-v3 and DenseNet-169.

Reference :