การจำแนกชนิดและนับไข่ไหมจากภาพโดยใช้โครงข่ายประสาทเทียม

Abstract

The Queen Sirikit Department of Sericulture has an important role in the silkworm industry by producing a disease free laying (DFL) for breeder. This task is also for the promotion and conservation of silkworm genetics. One of the main task is to count and classify eggs on the DFL. However, the traditional method for the classification and counting processes are performed by either humans or the estimation from the average weight of eggs. Those methods can be problematic and erroneous. Hence in the research, we use the counting neural networks (countNN) scheme in machine learning technology to perform the task instead. The input of the system are the scanned images of DFLs. The stochastic gradient descent (SGD) and Firefly algorithm (FA) is used to train the countNN. The class-wise counting result of the train data set from the SGD are 75.07%, 73.05%, 76.39% and 13.62% for the fresh eggs, all-blue eggs, shell eggs and dead eggs, respectively. The test result from the SGD system is 58.50% on the average. Whereas, that of the train data set from the FA are 89.24%, 84.14%, 88.01% and 77.14% for the fresh eggs, all-blue eggs, shell eggs and dead eggs, respectively. For the unfertilized egg, the FA system yields 36.21% and 35.73% for the all-blue and shell eggs stage. This is because of the similarity between all-blue in early stage and unfertilized eggs, as well as between shell eggs and unfertilized eggs. Finally, the overall counting rate from the FA system is 85.41%.