An explainable deep vision system for animal classification and detection in trail-camera images with automatic post-deployment retraining

This paper introduces an automated vision system for animal detection in trail-camera images taken from a field under the administration of Texas Parks and Wildlife Department. As traditional wildlife counting techniques are intrusive labor intensive to conduct, imaging is comparatively non-intrusive method capturing activity. However, given large volume produced trail-cameras, manual analysis remains time-consuming inefficient. We implemented two-stage deep convolutional neural network pipeline find animal-containing first stage then process these detect birds second stage. The classification classifies with overall 93% sensitivity 96% specificity. bird achieves better than sensitivity, 92% specificity, 68% average Intersection-over-Union rate. entire processes image less 0.5 seconds as opposed 30 human labeler. also addressed post-deployment issues related data drift features vary seasonal changes. utilizes automatic retraining algorithm update system. introduce novel technique detecting drifted triggering procedure. Two statistical experiments presented explain prediction behavior These investigate cues that steers towards particular decision. Statistical hypothesis testing demonstrates presence input significantly contributes system's decisions.