AUTHOR=Wang Fenmei , Liu Liu , Dong Shifeng , Wu Suqin , Huang Ziliang , Hu Haiying , Du Jianming 

TITLE=ASP-Det: Toward Appearance-Similar Light-Trap Agricultural Pest Detection and Recognition

JOURNAL=Frontiers in Plant Science

VOLUME=Volume 13 - 2022

YEAR=2022

URL=https://www.frontiersin.org/journals/plant-science/articles/10.3389/fpls.2022.864045

DOI=10.3389/fpls.2022.864045

ISSN=1664-462X

ABSTRACT=Automatic pest detection and recognition using computer vision techniques are a hot topic in modern intelligentagriculture, but suffer from a serious challenge: difficulty distinguishing the targets of similar pests in 2D images.The appearance-similarity problem could be summarized into two aspects: texture similarity and scale similarity.In  this  paper,  we  re-consider  the  pest  similarity  problem  and  state  a  new  task  for  the  specific  agriculturalpest  detection,  namelyAppearanceSimilarityPestDetection  (ASPD)  task.   Specifically,  we  propose  twonovel metrics to quantitatively define the texture-similarity and scale-similarity problems, in which the formeris  measured  by  Multi-Texton  Histogram  (MTH)  and  the  latter  is  measured  by  Object  Relative  Size  (ORS).Following  the  new  definition  of  ASPD,  we  build  a  task-specific  dataset,  named  PestNet-AS  that  is  collectedand re-annotated from PestNet dataset, and also a corresponding method ASP-Det.  In detail, our ASP-Det isdesigned to solve the texture-similarity by proposing a Pairwise Self-Attention (PSA) mechanism as well as Non-Local Modules to construct a domain adaptive balanced feature module that could provide high-quality featuredescriptors for accurate pest classification.  On the other hand, we also present a Skip-Calibrated Convolution(SCC) module that can balance the scale variation among the pest objects and re-calibrate the feature mapsinto the sizing equivalent of pests.  Finally, ASP-Det integrates the PSA-Non Local and SCC modules into a one-stage anchor-free detection framework with a center-ness localization mechanism.  Experiments on PestNet-ASshow that our ASP-Det could serve as a strong baseline for the ASPD task.