Research On Camouflage Animal Object Detection Based On Deep Learning

Camouflage is an important skill for animals to protect themselves in nature.It can help some animals integrate into the surrounding environment to avoid predators.Camouflage animal object detection has very important application value,which can be applied to the field of agriculture and natural ecology.The existing object detection algorithms have the following two difficulties in detecting such tasks.First,the camouflaged animals are extremely similar to the surrounding environment.The existing object detection algorithms are not prominent in the feature extraction ability of camouflaged animals,which reduces the detection performance.Second,overlapping occlusion is more common,which increases the difficulty of neural network training.Aiming at the above problems,CSN-YOLO high-precision camouflage animal detection algorithm and GDE-YOLO lightweight camouflage animal detection algorithm are proposed to adapt to camouflage animal object detection tasks.The main innovations are as follows :1.CSN-YOLO high-precision camouflage animal detection algorithmConv Next is used as a new backbone network to enhance the ability of network learning features and obtain richer semantic information.SPD-Conv is used for feature extraction and downsampling to reduce the performance degradation of the object detector in tasks with high similarity between foreground and background and low image resolution.After the feature extraction module,normalized attention is embedded to integrate channel information and spatial information,which improves network performance without increasing the number of parameters.2.GDE-YOLO lightweight camouflage animal detection algorithmGhost Net is designed as a new backbone network to speed up inference,reduce training resource consumption,and optimize inference performance.DO-Conv is used to save training memory resource overhead,keep the amount of calculation unchanged,and improve the performance of neural network.The ECA module is inserted after the C3 module of the 9th layer and Neck part of the network to enhance the feature information and global information representation ability and expand the receptive field.In order to verify the performance of the improved model in the camouflaged animal detection task,20 classes of target objects are extracted from the COD10 K dataset and data augmentation is performed to produce a dataset adapted to the current task.Ablation experiments and comparison experiments were conducted on this dataset separately to verify the effectiveness of CSN-YOLO algorithm and GDE-YOLO algorithm.The experimental results show that the CSN-YOLO algorithm has a m AP value of 93.7% and a recall rate of 90.4%,which have high accuracy and strong generalization ability in the task of camouflaged animal detection.With a m AP value of 93.4% and a speed of 72 frames per second,the GDE-YOLO algorithm achieves model lightness while maintaining high accuracy.