Weed infestations in crop fields remain one of the prime causes of reduction in agricultural productivity, thereby pushing for some intelligent solutions relevant to precision agriculture. With the help of deep learning (DL) and particularly Convolutional Neural Network (CNN), some object recognition methods can bring in high precision for discriminating between weed and crop. The performance evaluation of five CNN architectures, i.e., VGG16, MobileNetV2, EfficientNetB0, ResNet50, DenseNet121, is carried out by using a custom-collected dataset consisting of 2190 images containing annotations related to wheat field vegetation. A comprehensive comparative analysis of these five networks, relating to classification accuracy, loss, recall, and F1-score, is presented. It is observed from the study that CNN-based models can reliably and efficiently detect weeds and can therefore provide an impetus to researchers and practitioners in working towards deep-learning-based autonomous crop management systems.