An Efficient Pelican optimization based CNN-BiLSTM to Detect and Classify 3D Objects

Abstract

The lack of appropriate shape representation makes it complex to recognize the 3D shapes accurately and it is a hot topic in the field of Computer Vision (CV). This paper presents a Pelican optimized Convolutional Neural Network (CNN)-Bidirectional Long Short Term Memory (BiLSTM) to recognize the different objects in a particular scene. The CNN-BilSTM architecture is formed by placing two BiLSTM architectures below the CNN network and integrating the outputs via a fully connected layer. The pelican optimization algorithm is mainly incorporated to optimize the different hyperparameters associated with the CNN-BiLSTM architecture such as number of layers, batch size, number of layers, dropout, etc. The experiments are conducted using the ScanNet dataset which comprises both 2D and 3D data along with the labeled voxels. The proposed methodology offers improved results when compared with the existing techniques in terms of confusion matrix, accuracy, precision, and recall.