NEURAL NETWORKS FOR OBJECT DETECTION FROM ALL-SOURCE IMAGERY

LNK PROPOSES TO STUDY THE APPLICATION OF NEURAL NETWORKS TO OBJECT DETECTION FROM ALL-SOURCE IMAGERY. IN PHASE I EXISTING ARTIFICIAL NEURAL NETWORKS (ANNS) WILL BE EVALUATED ON OBJECT DETECTION PROBLEMS. THE STRENGTHS AND SHORTCOMINGS OF THESE TECHNIQUES WILL BE ASSESED AND NEW ANNS WILL BE DEVELOPED AS NECESSARY. AMONG THE TYPES OF NEW ANNS TO BE CONIDERED ARE HYBRID MODELS, WHICH MAY CONTAIN SEVERAL EXISTING OR NEW TYPES OF NEURAL NETWORKS, LINKED TOGETHER TO ENABLE SPECIFIC TYPES OF ANNS TO BE TAILORED TO PARTS OF A PROBLEM FOR WHICH THEY ARE MOST SUITABLE. IN ORDER TO USE IMAGERY FROM MULTIPLE SOURCES, THE IMAGERY MUST BE REGISTERED. LNK HAS DEVELOPED REGISTRATION TECHNIQUES FOR MATCHING IMAGES FROM DISSIMILAR SENSORS AND FOR MATCHING IMAGES AND MAPS. THESE TECHNIQUE WILL BE EMPLOYED AS NECESARY IN THE SYSTEM. IN PHASE II A USEABLE SYSTEM FOR OBJECT DETECTION WILL BE DEVELOPED USING THE BEST MODEL OR MODELS INVESTIGATED IN PHASE I.