Modified Deep Neural Network for Object Recognition

Abstract

Abstract: Object recognition has gained significance due to the rise in CCTV surveillance and the need for automated detection of objects or activities in images and videos. Lightweight process frameworks are in demand for sensor networks. While Convolutional Neural Networks (CNNs) are widely used in computer vision, many existing architectures are specialized. This paper introduces the Dimension-Based Generic Convolution Block (DBGC), enhancing CNNs with dimension-wise selection of kernels for improved performance. The DBGC offers flexibility for height, width, and depth kernels and can be applied to different dimension combinations. A key feature is the dimension selector block. Unoptimized kernel dimensions reduce computational operations and accuracy, while semi-optimized ones maintain accuracy with fewer operations. Optimized dimensions provide 5-6% higher accuracy and reduced operations. This work addresses the challenge of generic architecture in object recognition research.