A Conceptual Research Basedon Image Classification Based On Neural Network Architecture

Abstract

we propose a novel technique leveraging deep convolutional neural networks (DCNNs) in conjunction with scale-invariant feature transform (SIFT). Initially, an enhanced saliency method is employed to identify key points, from which point features are extracted. Subsequently, features are extracted from two deep CNN models—VGG and AlexNet. Following this, entropy-controlled method is applied to the DCNN pooling and SIFT point matrix to discern robust features.

The identified robust features are then fused into a matrix using a sequential approach, which is subsequently inputted into an ensemble classifier for recognition. This approach amalgamates the strengths of DCNNs and SIFT, leveraging their complementary capabilities to enhance classification accuracy, particularly in scenarios characterized by complex backgrounds, congestion, and object similarity.