CNN-based small object detection and visualization with feature activation mapping
Medhani Menikdiwela, Chuong Vinh Nguyen, Hongdong Li, Marnie Evelyn Shaw · 2017
Object detection is a well-studied topic, however detection of small objects still lacks attention. Detecting small objects has been difficult due to small sizes, occlusion and complex backgrounds. Small objects detection is important in a number of applications including detection of small insects. One application is spider detection and removal. Spiders are frequently found on grapes and broccolis sold at supermarkets and this poses a significant safety issue and generates negative publicity for the industry. In this paper, we present a fine-tuned VGG16 network for detection of small objects such as spiders. Furthermore, we introduce a simple technique called “feature activation mapping” for object visualization from VGG16 feature maps. The testing accuracy of our network on tiny spiders with various backgrounds is 84%, as compared to 72% using fined-tuned Faster R-CNN and 95.32% using CAM. Even though our feature activation mapping technique has a mid-range of test accuracy, it provides more detailed shape and size of spiders than using CAM which is important for the application area. A data set for spider detection is made available online.