Vision based automatic inspection of insects in pheromone traps

Abstract

Insects are one of the most important factors that threaten the yield efficiency in agricultural areas. Expenditures made for biological pesticides form a huge portion of the total expenses since insects massively reproduce. Observing the reproduction stages of the insects, more effective and smarter pesticizing scenarios can be achieved using biotechnical approaches such as pheromone traps rather than biological ones. Using pheromone traps, the massive reproduction is prevented since the male insects are attracted to the traps and cannot mate with the female ones. The most important disadvantage of the pheromone traps is the expensive labor cost due to the physical patrolling of the traps. Inspection of traps require expert staff who can recognize different kinds of insects. Besides the high labor costs, because of the human factor in the whole cycle, many problems occur such as errors made in counting and recording of the collected data. To overcome these problems, it is possible to integrate camera to the traps in order to lower the labor costs and assure more accurate record of the insect counts and types. Hence the visual data acquired through the traps can be inspected automatically using state of art computer vision techniques. The objective of this paper is to analyze and advance the methods that can discriminate and classify the insects in the traps under challenging illumination and environmental conditions using computer vision and machine learning algorithms. In this study, we use background subtraction and active contour models successively to separate the insects from the background and extract the outer boundary of the insects. We extract features using Hu moments (Hu), Elliptic Fourier Descriptors (EFD), Radial Distance Functions (RDF) and Local Binary Patterns (LBP). LBP features seem to outperform the rest of the features in recognition rate based on the individual performance of each method. The results from the underlying features are then fused using weighted majority voting to obtain a decision.