Individual Tree Crown Detection using GAN and RetinaNet on Tropical Forest

The detection performance of tree crowns in forest environment has not been satisfactory compared to common objects, especially using aerial RGB imagery. Previous methods regarding Individual Tree Crown Detection (ITCD) utilizes different data sources to improve the detection rate due to the noisy image. Image enhancement methods such as super-resolution provide a solution to the noisy image by reconstructing the image using the low-resolution image. Generative Adversarial Network (GAN)-based model has shown success in super-resolution techniques. However, the GAN-based model created artefacts that may hinder the accuracy of the detection. In this paper, a noise-cancelling GAN-based model is proposed by averaging the weights of a compressed image and non-compressed image. The proposed method forces the network to discriminate the noise to generate a more photorealistic image. This method is inspired by super-resolution GAN (SRGAN) architecture with Residual Dense Network as the generator network. A two-stage object detection RetinaNet model is then used to detect the individual tree crowns in a sequential fashion. Extensive experiments have been conducted on a self-assembled tree crown dataset which showed the proposed model is more superior than a non-enhanced model with 0.6017 and 0.5908 respectively. Based on the results of the proposed method, the super-resolution technique can be used in conjunction with object detection algorithm to improve the detection in ITCD to improve the detection rate.