11:00 〜 13:00
[MTT46-P07] Channel Attention-GAN based Synthetic Weed Generation for precise Weed identification
キーワード：生成モデル、データ収集
Abstract: Weeds are considered the main biological causative factor in declining crop yield and quality through competing with crops for sunlight, air, and other nutrient resources. So far, large-scale herbicide dispersal remains the primary means of weed control. However, due to the rise of herbicide resistance and the uneven distribution of weeds, under-dosing and over-dosing of herbicides are inevitable over the whole farmland. Therefore, the heavy use of herbicides does not guarantee the effectiveness of weed control, while not only bringing economic burden and leading to a series of environmental pollution problems such as reducing ecological diversity.
In order to realize the metaverse in agriculture, precise weed identification using artificial intelligence by implementing deep learning is the best solution in the context of big data. These techniques usually need very large datasets coping with real-world conditions, however, cultivating and labeling weeds manually are very specialized and challenging tasks, requiring specialized knowledge and skills in weed science and agronomy.
To address such data problems, our research proposed data farm which consists of 1) standardized weed data cultivation and collection scheme. 2) generative model-based synthetic data generation pipeline. Our data farm first collects data fully automatically and remotely through the data cultivation and collection scheme, then generates more diverse synthetic data by learning the latent representations of real-world data.
To be detailed, first, weed data cultivation and collection scheme was designed by combining farmbot (Farmbot Inc, San Luis Obispo, USA), NVIDIA Jetson Xavier NX (NVIDIA Corporation, Santa Clara, USA) and multiple ELP 13mp cameras (Ailipu Technology Co.,Ltd, Shenzheng, China). And an FTP server was built on Google Cloud Platform to achieve real-time synchronization cloud data. Tens of thousands of weed data can be collected within a month using our scheme.
Second, generative model-based synthetic data generation pipeline contains data cleansing and a GAN model driven by channel attention mechanism. To evaluate our proposed data generation pipeline, a metric called Frechet Inception Distance (FID) score was used to assess the quality of images created by the generative model (the lower the score, the more realistic the sample), the FID score of the proposed method was 25.67 that is better than current state-of-the-art network SNGAN scored 35.78 and wasserstein auxiliary classification GAN with gradient penalty (WACGAN-GP) scored 51.01. Besides, the accuracy of generating specified species of weed image was also evaluated using an external classifier, achieving an average identification accuracy of 91.02%, obtaining better generation results than SNGAN (90.94%), and WACGAN-GP (83.92%).
In order to realize the metaverse in agriculture, precise weed identification using artificial intelligence by implementing deep learning is the best solution in the context of big data. These techniques usually need very large datasets coping with real-world conditions, however, cultivating and labeling weeds manually are very specialized and challenging tasks, requiring specialized knowledge and skills in weed science and agronomy.
To address such data problems, our research proposed data farm which consists of 1) standardized weed data cultivation and collection scheme. 2) generative model-based synthetic data generation pipeline. Our data farm first collects data fully automatically and remotely through the data cultivation and collection scheme, then generates more diverse synthetic data by learning the latent representations of real-world data.
To be detailed, first, weed data cultivation and collection scheme was designed by combining farmbot (Farmbot Inc, San Luis Obispo, USA), NVIDIA Jetson Xavier NX (NVIDIA Corporation, Santa Clara, USA) and multiple ELP 13mp cameras (Ailipu Technology Co.,Ltd, Shenzheng, China). And an FTP server was built on Google Cloud Platform to achieve real-time synchronization cloud data. Tens of thousands of weed data can be collected within a month using our scheme.
Second, generative model-based synthetic data generation pipeline contains data cleansing and a GAN model driven by channel attention mechanism. To evaluate our proposed data generation pipeline, a metric called Frechet Inception Distance (FID) score was used to assess the quality of images created by the generative model (the lower the score, the more realistic the sample), the FID score of the proposed method was 25.67 that is better than current state-of-the-art network SNGAN scored 35.78 and wasserstein auxiliary classification GAN with gradient penalty (WACGAN-GP) scored 51.01. Besides, the accuracy of generating specified species of weed image was also evaluated using an external classifier, achieving an average identification accuracy of 91.02%, obtaining better generation results than SNGAN (90.94%), and WACGAN-GP (83.92%).