Improved Yolov7 Tiny with Global Attention Mechanism for Camouflage Object Detection

Detecting camouflaged objects presents a greater challenge in comparison to normal object detection as an object is highly similar to a complex background. Due to the similarities of an object with the background, it is very difficult to extract complete feature information of that target, causing an increase in false and missed detection. To analyse these issues, this research article introduces an enhanced algorithm built upon Yolov7 tiny. Firstly, a GAM was introduced in the backbone of Yolov7 tiny for the extraction of intricate features of an object. Also, three attention layers were introduced in the head section to increase prediction confidence. The parameters of the GAM module were adjusted and the position of the GAM was properly selected to ensure that the proposed method offered the best accuracy in camouflage detection. For training purposes, a Moving Camouflaged Animal dataset was considered which was initially available in video format and was later converted into images. It may be noted that seven classes have been considered to have a high camouflage environment. The problem of false and missed detection for camouflage images with complex dynamic backgrounds was improved giving better detection results by the introduced approach. The findings obtained from the experimentation revealed that the proposed model has attained an increase in mean Average Precision by 10.7% and 7.2% for mAP at 0.5 and mAP at 0.5:0.95 respectively for validation images for all classes and testing images there was an increase in mAP by 10.5% and 9.2% respectively.