Hunger games search based deep convolutional neural network for crop pest identification and classification with transfer learning

Agriculture is considered the backbone for developing a country both economically and financially. The quality of food crops is diminished only due to the attack of crop pests. So, pest attacks must be considered the main cause in the agriculture sector. Machine learning models have been introduced to overcome the problem of classification and detection of pests and attain the best solution. But, the technology shows poor performance in classifying and detecting insects with similar feature types and different positions in the natural environment. So in this research, an optimized deep learning model is proposed for efficient pest detection with better accuracy. This study proposed a hunger games search-based deep convolutional neural network (HGS-DCNN) model to classify the crop pest images. Therefore, in this research, a new convolutional layer is introduced to reduce the parameter redundancy in the CNN model. The research is processed in two stages: pre-processing and augmentation and pest classification. The pre-processing stage is employed with a new adaptive cascaded filter (ACF) to enhance the visual appearance and quality of an image. The proposed filtering model is cascaded with decision-based median filtering (DMF) and the guided image filtering (GIF) approach. The highly discriminative features are extracted in the classification stage, and the field crop pest images are classified very efficiently. The proposed pest classification model was evaluated with pre-trained learning architectures such as ResNet50, Efficient Net, Dense Net, Inceptionv3 and VGG-16 pest classification models. The proposed model procures an accuracy, precision, F-score, sensitivity and specificity of 99.1, 97.80, 97.80, 97.82 and 99.43%, respectively. The K-fold cross-validation and ablation study is conducted in this research to prove the model’s efficacy. Also, the effectiveness of the proposed model is validated with the hyper-parameters such as learning rate, the number of epochs and mini-batch size consecutively.