MAM-IncNet: an end-to-end deep learning detector for Camellia pest recognition

Camellia oil is one of the most healthy edible oils in the world. It has the effects of lowering blood pressure, reducing blood fat, and softening blood vessels. Whereas, the Camellia oleifera plant is easily infected by various pests and diseases in the process of growing, which limits the yield of Camellia oil. Thereupon, seeking an intelligent information tool to automatically detect Camellia oleifera pests is of great importance. Recent development in deep learning (DL)-based methods has provided promising performance in plant pest detection. However, to date, DL methods have been rarely applied in this field, except that some existing work focuses on it from public datasets. The main reasons behind the limited usage of DL models in Camellia pest detection include: a large number of training samples, which is difficult to collect, the complicated backdrops of experimental materials, which are not easy to train an efficient model, the low recognition accuracy, which is hard to apply in practical scenarios, and others. Therefore, this study proposes a novel network architecture, namely MAM-IncNet, to address these challenges. Referring to the cascaded structure of single shot multibox detector (SSD), we substitute the former convolutional layers of SSD with the optimized Inception modules (M-Inception), and the pre-trained VGG16 is utilized as the backbone network. Further, a hybrid attention mechanism including channel-wise and spatial attention is incorporated into the network to realize the maximum reuse of inter-channel relationships and spatial point characteristics. The proposed method has attained a recall rate of 81.44% for the detection of Camellia oleifera pests in practical field scenarios. Experimental findings demonstrate the efficacy and feasibility of the proposed method for the detection of Camellia oleifera insect pests.