Agricultural light-trapped pest detection methods under unbalanced data

Abstract

This paper designs an improved YOLOX method for agricultural pest detection under unbalanced data. With the expansion of food demand due to rapid population growth in recent years, many crops have been damaged by pests due to frequent natural disasters, which have caused serious damage to farmers’ economic development. As an important biological disaster in agricultural production in China and the world, crop pests and diseases are the biggest cause of sustainable and stable development of agricultural production, with a wide range of species, high impact and high potential for outbreaks becoming its label. Accurate detection of pests is therefore an urgent necessity for the excellent development of the crop industry. In this paper, the binary cross-entropy loss in YOLOX is changed to Focal loss, an attention mechanism is added to its backbone feature extraction network Backbone to make the model more edge-oriented, and a depth-separable convolution is introduced to reduce the number of parameters. The improved pest detection model obtained a recall of 93% and an average accuracy of 87.6%, which can be effectively applied in real life.