(1.The Colleges and Universities of Jiangxi Province for Key Laboratory of Information Technology in Agriculture and Software Institute， Jiangxi Agricultural University， Nanchang 330045， China； 2.College of Software， Jiangxi Agricultural University， Nanchang 330045， China； 3.College of Computer and Information Engineering， Jiangxi Agricultural University， Nanchang 330045， China； 4.University of Debrecen，Debrecen 4032，Hungary)
Abstract： Fruit fly is a kind of quarantine pest that attracts much attention at home and abroad. There are many kinds of fruit flies. Different kinds of fruit flies are similar in shape and size, which is difficult to identify. In addition, in practical applications, it is difficult to identify fruit flies due to the lack of information about shielding, view-point, changing light and shadow and other factors. This study proposes a bilinear pooled attention network for fruit fly classification to learn effective discriminant characteristics. The network is composed of two parts: saliency feature module and cross-layer bilinear feature module. Saliency feature module realizes feature enhancement by filtering enhancement processing of two different convolution layers. Cross-layer bilinear module is based on bilinear pooling fusion features, determines the attention location, and mines discriminant features. Experiments on fruit fly’s data set with natural environment background show that the method is effective and has good practical application prospect.
Key words : fruit fly detection；bilinear pooling；attention mechanism
