Lw-PSCNN: Lightweight Pointwise-Separable Convolution Neural Network for ISAR Image Classification

Inverse synthetic aperture radar (ISAR) is a remote-sensing technique used in several applications due to its advantage of improved resolution. The use of lightweight models for classification helps in terms of resource-efficient usage and real-time purposes, especially in portable devices such as handheld imaging systems, small drone imaging systems, etc. The remote-sensing classification models in the literature either have a very high number of parameters or perform poorly. This article proposes a model called lightweight pointwise-separable convolution neural network (Lw-PSCNN), considering classification accuracy and inference time as the main objectives. Two different ISAR datasets with different input resolutions are considered for performance evaluation. The proposed model has around 1.24 M parameters and achieved test accuracy of 99.3% and 98.4% and inference time of 6.649 and 6.779 ms for the two datasets, respectively, on the NVIDIA Jetson TX2 device. The experimental results show that the Lw-PSCNN model has produced better accuracy using very few trainable parameters and multiply accumulate operations, with very little inference time compared to the state-of-the-art remote sensing classification models.